cognitive abilities, personality and interests
TRANSCRIPT
This thesis has been submitted in fulfilment of the requirements for a postgraduate degree
(e.g. PhD, MPhil, DClinPsychol) at the University of Edinburgh. Please note the following
terms and conditions of use:
• This work is protected by copyright and other intellectual property rights, which are
retained by the thesis author, unless otherwise stated.
• A copy can be downloaded for personal non-commercial research or study, without
prior permission or charge.
• This thesis cannot be reproduced or quoted extensively from without first obtaining
permission in writing from the author.
• The content must not be changed in any way or sold commercially in any format or
medium without the formal permission of the author.
• When referring to this work, full bibliographic details including the author, title,
awarding institution and date of the thesis must be given.
Cognitive abilities, personality and interests:
Their interrelations and impact on occupation
Jason Timothy Major
Doctor of Philosophy in Psychology
The University of Edinburgh
2013
ii
Declaration
I hereby declare that this work has been composed by me, and that it is my own
work, except where it has been clearly indicated. Furthermore, the work has not
been submitted for any other degree or professional qualification.
Jason Major
iii
Acknowledgements
I would like to thank my supervisors, Dr. Wendy Johnson and Professor Ian Deary,
for their great expertise and encouragement. During the PhD I benefited from an
intellectually lively department, particularly in the area of Individual Differences. I
would like to thank all those students and staff who contributed to it. I also would
like to thank my parents and family for their support.
iv
Abstract
Cognitive ability, personality and interests are three distinct topics of investigation
for psychology. In the past two decades, however, there have been growing appeals
for research and theories that address the overlap among these domains (Ackerman
& Heggestad, 1997; Armstrong, Day, McVay, & Rounds, 2008). One example of
such a theory is PPIK theory (intelligence-as-process, personality, interests, and
intelligence-as-knowledge) by Ackerman (1996). Integrative theories have the
potential of not only increasing our theoretical understanding of the development of
these individual differences, but of and improving vocational guidance through better
prediction of future occupation (Armstrong, Su, & Rounds, 2011; Johnson &
Bouchard, 2009). The research of this thesis was centered on examining the links
among cognitive ability, personality and interests. The data came from Project
TALENT (PT), a nationally-representative sample of approximately 400,000
American high school students from 1960 (Flanagan et al., 1962). A secondary topic
was whether an integrated view could improve the prediction of attained occupation.
This was tested with occupational data from follow-up PT surveys, conducted 11
years after high school. The first study addressed the structure of the PT intelligence
tests. Three popular models of intelligence were compared through factor analysis:
the Extended Fluid-Crystallized (Gf-Gc), Cattell-Horn-Carroll (CHC) and Verbal-
Perceptual-Image Rotation (VPR) models. The VPR model provided the best fit to
the data. The second study was an investigation of linear and nonlinear intelligence-
personality associations in Project TALENT. The ten PT personality scales were
related to the Big Five personality factors through content examination, consistent
with previous research (Reeve, Meyer, & Bonaccio, 2006). Through literature
review of studies on intelligence and the Big Five, 17 hypotheses were made about
linear associations and quadratic associations of personality traits with general
intelligence (g). The majority of the hypotheses were supported in all four grade
samples: 53% in male samples, and 58% in female samples. The most notable
finding, contrary to previous research, was that quadratic associations explained
substantive variance above and beyond linear effects for Sociability, Maturity, Vigor
and Leadership in males, and Sociability, Maturity and Tidiness in females. The
third study examined associations between cognitive ability and interests, and their
v
capacity to predict occupational type. Specifically, Ackerman’s PPIK theory
suggests that there are two “trait complexes” that are combinations of cognitive
abilities and interests (termed science/math and intellectual/cultural). Trait
complexes were derived from PT data separately by latent class analysis and factor
analysis. It was hypothesized that they should have validity equal to or greater than
individual intelligence and interests scores in predicting attained occupation.
Instead, trait complexes derived through latent class analysis predicted substantially
less variance in occupation than individual scales. The factor-analytic trait
complexes performed more like the scales, but one trait complex (which involved g
centrally) was inconsistent with PPIK theory. Overall, the trait complexes of PPIK
theory were not supported. The results of the three studies are discussed in the
context of existing integrative theories, and suggestions for future research are
provided.
vi
Publications from this thesis
Major, J. T., Johnson, W., & Deary, I. J. (2012). Comparing models of intelligence
in Project TALENT: The VPR model fits better than the CHC and Extended Gf-Gc
models. Intelligence, 40(6), 543-559.
Major, J. T., Johnson, W., & Deary, I. J. (in press). Linear and nonlinear relations
between personality and general intelligence in Project TALENT. Journal of
Personality and Social Psychology.
Major, J. T., Johnson, W., & Deary, I. J. (2013). Trait complexes of cognitive
abilities and interests and their relations to realized occupation. Manuscript in
preparation.
vii
Table of contents (text)
Acknowledgements .................................................................................................... iii
Abstract ...................................................................................................................... iv
Publications from this thesis .................................................................................... vi
Table of contents ...................................................................................................... vii
Chapter 1: Introduction ............................................................................................ 1
1.1 Theories of intelligence, personality and interests ............................................ 2
1.1.1 Intelligence .................................................................................................. 2
1.1.2 Personality ................................................................................................... 4
1.1.3 Occupational interests .................................................................................. 6
1.3 Integrative theories ............................................................................................ 7
1.3 Prediction of occupational type ....................................................................... 12
Chapter 2: Project TALENT’s design and measures ........................................... 14
2.1 Intelligence tests ............................................................................................... 15
2.2 Personality tests ................................................................................................ 16
2.3 Occupational interest tests ................................................................................ 17
2.4 Occupation at follow-up ................................................................................... 18
Chapter 3: Comparing the VPR, CHC and Extended Gf-Gc models ................. 20
3.1 Introduction ...................................................................................................... 20
3.1.1 Previous factor-analytic research on Project TALENT ............................. 26
3.2 Methods ............................................................................................................ 29
3.2.1 Sample ....................................................................................................... 29
3.2.2 Measures .................................................................................................... 29
3.2.3 Data preparation ......................................................................................... 32
3.2.4 Analysis method ........................................................................................ 34
3.3 Results .............................................................................................................. 35
3.3.1 Exploratory factor analysis ........................................................................ 35
3.3.2 Confirmatory factor analyses ..................................................................... 41
3.3 Discussion ........................................................................................................ 52
3.3.1 The three theories in Project TALENT ..................................................... 53
3.3.2 Variations in VPR model specifications .................................................... 56
3.3.3 Theoretical implications for the structure of intelligence .......................... 57
3.4 Linking cognitive ability with personality ....................................................... 60
Chapter 4: Linear and nonlinear associations between general intelligence and
personality ................................................................................................................. 62
4.1 Introduction ...................................................................................................... 62
4.1.2 Linear personality-intelligence associations in Project TALENT ............. 68
4.1.3 Possible nonlinear associations .................................................................. 69
viii
4.2 Method .............................................................................................................. 71
4.2.1 Sample ....................................................................................................... 71
4.2.2 Intelligence measures ................................................................................. 71
4.2.3 Personality measures ................................................................................. 72
4.2.4 The general factor of personality ............................................................... 72
4.2.5 Methods of analysis ................................................................................... 77
4.3 Results ........................................................................................................... 78
4.3.1 LMS results compared to GAM results ..................................................... 81
4.3.2 Grade and sex differences .......................................................................... 82
4.3.3 Figures 4.1 to 4.3: titles and captions ........................................................ 82
4.4 Discussion ........................................................................................................ 87
4.4.1 Conclusions and future directions ............................................................. 94
4.5 Integrating cognitive abilities and interests ...................................................... 95
Chapter 5: Trait complexes of cognitive abilities and interests and their
predictive validity for occupation ........................................................................... 96
5.1 Introduction ...................................................................................................... 96
5.1.2 Previous Project TALENT research ........................................................ 101
5.2 Method ............................................................................................................ 103
5.2.1 Sample ..................................................................................................... 103
5.2.2 Intelligence measures ............................................................................... 104
5.2.3 Interest measures ..................................................................................... 105
5.2.4 Occupational categories ........................................................................... 106
5.2.5 Method of analysis ................................................................................... 106
5.2.6 Interest and cognitive ability factors ........................................................ 107
5.3 Results ............................................................................................................ 111
5.3.1 Factor-analytic trait complexes ................................................................ 111
5.3.2 Latent class trait complexes ..................................................................... 116
5.3.3 Prediction of occupational type ............................................................... 118
5.3.4 Multinomial prediction ............................................................................ 124
5.4 Discussion ...................................................................................................... 125
Chapter 6: Conclusion ........................................................................................... 134
5.1. Cognitive ability ............................................................................................ 134
5.2. Personality-intelligence associations ............................................................. 134
5.3. Trait complexes ............................................................................................. 137
5.4. Suggestions for future research ..................................................................... 140
References ............................................................................................................... 142
Appendix A ............................................................................................................. 153
Appendix B ............................................................................................................. 155
Appendix C ............................................................................................................. 158
ix
Table of contents (Figures and Tables)
Figure 1.1 Holland’s interest hexagon 6
Table 3.1 Primary features of the CHC, Extended Gf-Gc and VPR models 22
Table 3.2 Project Talent test names, short descriptions, and reliabilities for
males/females 31
Table 3.3 Factor pattern matrices for grade 10 males/females in the broad selection
of PT tests 37
Table 3.4 Factor pattern matrices for grade 10 males/females in the narrow selection
of PT tests 40
Table 3.5 Factor correlation matrices for grade 10 males (below diagonal) and females
(above diagonal) in the broad and narrow selections of PT tests 41
Table 3.6 First-order loadings for the CHC, Extended Gf-Gc and VPR models in
the broad selection (grade 10 males) 43
Figure 3.1 Measurement model of the VPR model with factor loadings from
the grade 10 male sample 45
Figure 3.2 Measurement model of the Extended Fluid-Crystallized model with factor
loadings from the grade 10 male sample 46
Figure 3.3 Measurement model of the Cattell-Horn-Carroll model with factor loadings
from the grade 10 male sample 46
Table 3.7 Fit statistics of confirmatory factor models for the broad selection of PT tests 49
Table 3.8 First-order loadings for the CHC, Extended Gf-Gc and VPR models
in the narrow selection (grade 10 males) 51
Table 3.9 Fit statistics of confirmatory factor models for the narrow selection
of PT tests 52
Table 4.1 Associations of the Project TALENT personality scales with the Big Five 65
Table 4.2 Personality test descriptives 72
Table 4.3 Correlations among personality scales after removal of the general personality
factor (frade 10 males/females) 76
Table 4.4 Standardized linear and quadratic effects of g on the personality scales (males) 79
Table 4.5 Standardized linear and quadratic effects of g on the personality scales (fem.) 80
Figure 4.1 Mean personality as predicted by general intelligence (grade 10 males) 84
Figure 4.2 Mean personality as predicted by general intelligence (grade 10 females) 85
Figure 4.3 LMS and GAM-predicted sociability as a function of general intelligence
(grade 10 males) 86
Table 5.1 Occupation categories and sample percentages (grade 12 sample) 106
Table 5.2 Factor loadings for grade 12 males/females in the confirmatory
intelligence model 110
x
Table 5.3 Correlations matrix for interest composites and cognitive ability factors
(grade 12 males/females) 112
Table 5.4 CFA solution of interests and abilities (grade 12 males) 114
Table 5.5 CFA solution of interests and abilities (grade 12 females) 115
Table 5.6 Latent class means from LCA (grade 12 males) 117
Table 5.7 Latent class means from LCA (grade 12 females) 118
Table 5.8 Odds ratios of abilities and interests predicting job categories
(grade 12 males) 120
Table 5.9 Odds ratios of abilities and interests predicting job categories
(grade 12 females) 121
Table 5.10 Odds ratios of latent classes in predicting job category (grade 12 males) 123
Table 5.11 Odds ratios of latent classes in predicting job category (grade 12 fem.) 123
Table 5.12 Original sample composition and correct classification percentages for
multinomial regression (grade 12 males) 125
Table 5.13 Original sample composition and correct classification percentages for
multinomial regression (grade 12 females) 125
1
Chapter 1: Introduction
The two main domains of study in differential psychology are intelligence
and personality. More recently, occupational interests have received increased
attention, driven by the practical goal of improving vocational guidance. These three
domains of individual differences are not entirely independent, but overlap
(Ackerman & Heggestad, 1997; Barrick, Mount, & Gupta, 2003). This overlap has
led researchers to call for an integrative theory of individual differences that takes
intelligence, personality and interests into account (Ackerman & Heggestad, 1997;
Armstrong et al., 2008).
The overall purpose of this thesis was to explore the relations among the
intelligence, interests and personality domains. The secondary goal was to examine
the potential for an integrated view to improve the prediction of attained occupation.
In this introductory chapter, a review is provided of prominent theoretical models in
the three domains. Research on their integration is then reviewed, and finally the
ability of integrative theories to improve our understanding of occupational
attainment.
The data used in this thesis came from Project TALENT (PT), a longitudinal and
nationally-representative study of the aptitudes, interests, and backgrounds of
American high school students, started in 1960. Chapter 1 provides background on
how PT was conducted and the measures within it. Chapter 2 is a study that
examined the structure of the PT intelligence tests, comparing three of the
predominant intelligence models in the literature. Chapter 3 is a study that examined
linear and non-linear associations between personality and intelligence. Chapter 4 is
a study that focused on the associations between intelligence and occupational
interests. It examined the character of potential “trait complexes” of intelligence and
interests in the PT scales, and whether these trait complexes are better or worse
predictors of future occupational type than individual scores for cognitive ability and
interests. In chapter 5, a summary is given of what these studies have revealed about
the intersection of intelligence, personality and interests, and the potential of
2
integrative frameworks both to describe this overlap and be useful in the prediction
of occupation.
1.1 Theories of intelligence, personality and interests
In order to study individual difference variables, psychologists need not only
valid and reliable measures, but for multifaceted traits such as intelligence,
personality and interests, theoretical models of their makeup. In this section, a brief
but up-to-date picture is provided of the main theories for the three domains.
1.1.1 Intelligence
Research on the structure of intelligence has continued uninterrupted since
the early twentieth century, when Spearman (1904) first proposed the concept of
general intelligence (g), as the common factor underlying all cognitive ability tests.
In the past several decades, research has converged on the hierarchical model as the
best representation of cognitive abilities (Hunt, 2011; Reeve & Bonaccio, 2011).
Here the term ‘hierarchical’ is used in the sense of a multiple-strata model, in which
higher-order or more general cognitive ability factors are proposed to contribute
directly to the lower-order or more specific ability factors. The lowest factor stratum
consists of narrow abilities measured by individual tests. The second stratum
consists of broad ability factors that emerge from higher-order factor analysis of
narrow abilities. The third stratum emerges from factor analysis of broad abilities,
but at this level only a single factor, known as g, is typically found.
Hierarchical models have both advantages and disadvantages in describing
intelligence. One advantage is that the highly general concept of intelligence is
divided into more manageable components called cognitive abilities (Reeve &
Bonaccio, 2011). A cognitive ability can be defined as a latent trait that is observed
from performance on particular cognitive tasks. Each ability is assessed by multiple
tests, which vary in how purely they tap the ability (the remainder of test variance is
made up specific test variance and cross-loadings on other factors that ideally are
small in magnitude). A disadvantage of these models is that they are dependent to a
certain extent on the properties of the tests in the battery, and on the testing sample
(Hunt, 2011). In addition, subjectivity remains in interpreting to what the factors
3
correspond at a more basic level, such as in a cognitive or biological sense (Hunt,
2010). An ultimate purpose of structural theories is to provide precise enough
delineations of factors that their biological bases can be discovered, although this
remains largely a future goal. Nonetheless, structural models have contributed
greatly to advances in intelligence research, and are an essential part of current
theories.
The three best-supported models in the intelligence literature are the extended
fluid-crystallized (Gf-Gc) model (Horn & Blankson, 2005), the Cattell-Horn-Carroll
model (CHC) model (McGrew, 2005, 2009), and the verbal-perceptual-image
rotation (VPR) model (Johnson & Bouchard, 2005b). The differences among these
models are covered in greater detail in chapter three, and are only summarized here.
The models diverge primarily at the second stratum of the intelligence hierarchy. 1
The CHC model contains the greatest number of second-order factors: ten that have
been firmly identified, and six more that have been characterized as “tentative”
(McGrew, 2009). The extended Gf-Gc model contains eight second-stratum factors,
which overlap strongly with those in the CHC model. This reflects the common
origin of the two models, which can be traced back to Cattell’s original fluid-
crystallized model (Cattell, 1963), and Thurstone’s primary mental abilities
(Thurstone, 1938). The VPR model, by comparison, is more parsimonious and
proposes only three second-stratum factors (the factors for which it was named).
The most notable difference between the second-stratum factors in the three
models is that the CHC and Gf-Gc models contain factors which are delineated by
how much they tap so-called fluid versus crystallized ability. Fluid intelligence
refers to the ability to learn new information and solve novel problems, without
regard to knowledge content or the content of material to which reasoning is to be
applied, whereas crystallized intelligence refers to knowledge acquired from
previous learning experiences (Cattell, 1963). These two factors are both present in
the second-stratum of the CHC and Gf-Gc models. Moreover, the other factors can
1 There has been some confusion in the literature surrounding the term stratum. Typically, this term
has meant a level of a hierarchical model containing one or more factors. However, Johnson &
Bouchard (2005b) characterized the VPR model as having four strata, counting the first level of
individual tests as a stratum (p. 397). In the traditional sense, the model only has three strata. Reeve
and Bonaccio (2011) also inaccurately presented the VPR model as having four strata.
4
be distinguished into those based more on “process” (Gf) compared to those based on
“content” (Gc; Carroll, 1993). The VPR model, in contrast, posits that the second-
stratum factors are distinguished only by their content. The factors in the VPR
model are thought to be formed because the tests differ in the extent they are verbal
(requiring the understanding of words and symbols), perceptual (requiring the
understanding of visual-spatial stimuli), or image rotational (requiring the mentally
rotation of visual-spatial stimuli (Johnson & Bouchard, 2005b)
Three factor-comparison studies have compared the three models presented,
and in each case the VPR model displayed the best statistical fit (Johnson &
Bouchard, 2005a, 2005b; Johnson, Te Nijenhuis, & Bouchard, 2007). However,
these comparison studies relied on previous versions of the CHC model (the three-
stratum model; Carroll, 1993) and the extended Gf-Gc model (the Gf-Gc model
presented by Horn, 1998). Thus, a new comparison study was needed to distinguish
among the three models. This is presented in the third chapter. In addition, the best-
supported model was to be used in further examining associations with personality
and interests.
1.1.2 Personality
The dominant model in personality psychology is the Five-Factor Model
(FFM), which was first developed in factor analyses of personality trait terms by
Tupes and Christal (1961; reprinted in 1992) and Norman (1963). However, the
“Big Five” model only gained prominence in the mid-1980s after several different
researchers found new empirical support for the model and argued for its theoretical
merit (Goldberg, 1990; McCrae & Costa, 1985). Numerous labels have been put
forward for each of the factors; however, the most commonly-used names were
proposed by Costa and McCrae (1992): Extraversion, Neuroticism, Openness to
Experience, Agreeableness and Conscientiousness.
The trait approach to personality has itself undergone many criticisms since
its inception (Deary, 2009). Some have argued that the FFM is simply an empirical
taxonomy, and thus that it lacks theoretical explanations for what the personality
traits are, and how they emerge developmentally (Cervone, 2005; Cramer et al.,
2012). Notwithstanding these more basic issues surrounding traits, the FFM has
5
received considerable support as a taxonomic framework. The Five-Factor structure
has been studied and partially replicated in over fifty cultures (McCrae &
Terracciano, 2005, but see De Raad & Peabody, 2005). In addition, it has been
found to capture the variance of personality factors on other major scales, such as the
Eysenck Personality Questionnaire (Costa & McCrae, 1995), and Cattell’s 16PF
(Conn & Rieke, 1994, cited in McCrae, 2009).
There is the possibility that additional factors should be added to the Big
Five. For example, Ashton and Lee (2005) proposed a sixth dimension termed
Honesty-Humility, and up to eight broad factors have been supported by lexical
studies (De Raad & Barelds, 2008). Nonetheless, these studies have also recovered
the Big Five, supporting the position that they are “more-or-less sufficient to account
for the co-variation of most personality traits” (McCrae, 2009, p. 148). Cramer et al.
(2012) criticized the FFM for not being able to account for the variance in trait
ratings without cross-loadings; however, they did not specify a hierarchical model (in
the sense outlined above) that included facets. FFM proponents acknowledge that
facet-level variance is a significant part of the FFM model. As Ashton and Lee
(2012) observed: “Researchers have also known that a few broad factors can account
for some large fraction of the covariation among personality variables, and not for all
that covariation” (p. 433). Further refinement of the FFM at the facet and item level
is still ongoing (McCrae, 2009).
The PT personality scales were not developed according to the FFM, but as
described further in chapter 2, one study found a moderate level of correspondence
between the PT scales and the Big Five (Reeve, Meyer & Bonaccio, 2006). The
research on personality here was done in reference to the FFM, because it has proven
a useful taxonomy for personality psychology. Moreover, the FFM has been used in
much of the research aimed at discovering associations of personality with other
individual difference variables (Ackerman & Heggestad, 1997; Barrick et al., 2003),
and in the context of occupational prediction (Judge, Higgins, Thoresen, & Barrick,
1999). Thus the use of the FFM was helpful in forming hypotheses and relating the
results back to the literature.
6
1.1.3 Occupational interests
Similar to the situation in personality psychology, one model is predominant
in the occupational interests field: the RIASEC model (Holland, 1959, 1997). The
model is composed of six broad interest factors: Realistic, Investigative, Artistic,
Social, Enterprising, Conventional. The interest types are conceived as both
manifestations of different work environments, and people’s preferences for these
environments (Holland, 1959, 1997). The six types are organized in a hexagon, in
which the relations are expected to be highest between adjacent types, followed by
alternative types (types separated by one in the hexagon), whereas opposite types are
expected to have zero or negative associations. Consistent with these proposed
associations, Prediger (1982) analyzed Holland interest scores for career groups and
individuals and found support for two dimensions, named Data/Ideas and
People/Things. These dimensions spanned opposite types: People/Things contrasted
Social with Realistic interests, while Data/Things contrasted Conventional and
Enterprising on one side with Investigative and Artistic on the other. Figure 1.1
Displays the Holland hexagon and Prediger’s two dimensions. Hogan (1983) found
two similar dimensions, which he called sociability and conformity, although the
axes of these dimensions were rotated 30 degrees clockwise from Prediger’s
dimensions (Armstrong et al., 2011).
Figure 1.1 Holland’s interests hexagon
7
In addition to these two dimensions that appear to underlie the RIASEC
hexagon, a third dimension has been found termed prestige (Einarsdóttir & Rounds,
2000; Tracey & Rounds, 1996). This dimension was found when additional
occupations were added to the scale underlying the RIASEC typology: the vocational
preference inventory (VPI; Holland, 1985). These results suggested that the VPI has
a restricted range of occupational prestige, an observation that was confirmed when a
broader range of U.S. occupations was examined (Deng, Armstrong, & Rounds,
2007). In that same study, it was found that the prestige dimension was in fact not
orthogonal to People/Things and Data/Ideas. The prestige dimension was associated
with the Ideas pole of the Data/Ideas dimension. In addition, gender differences in
occupational preferences were strongly associated with the People/Things dimension.
Men tended to prefer jobs towards the Things pole, while women tended to prefer
jobs towards the People pole, as had been observed in previous studies (Tracey &
Rounds, 1992). Thus, the dimensions that underlie the RIASEC hexagon are related
to prestige and gender differences, but the prestige dimension is only observed if a
wider range of occupations than on the VPI is used (Deng et al., 2007).
To summarize, it appears that the RIASEC typology with its two dimensions
is a reasonably adequate description of occupational interests for occupations, but
this description could and probably should be expanded to give greater emphasis to a
third dimension of prestige. The RIASEC model of interests is the primary one used
in integrative research (Ackerman & Heggestad, 1997; Armstrong et al., 2008); thus
it was an important point of reference for the research presented here.
1.3 Integrative theories
This section focuses on integrative theories in general, and does not cover all the
literature on the overlap between interests, personality and cognitive abilities. For a
review of cognitive ability and personality, see chapter 4. For a review of cognitive
ability and interests, see chapter 5. The overlap of personality and interests was not
reviewed because it was not addressed in the research here. There were two primary
reasons for this. First, I suspected that cognitive abilities are the primary drivers of
associations and so wanted to focus first on their associations. Second, it was
8
necessary to keep the number of analyses in the study manageable (see chapter 5 for
further discussion).
Psychologists have long hypothesized and observed that cognitive abilities,
personality and interests are not entirely independent, but related; studies were
conducted as early as Pearson in 1906 (see Ackerman & Heggestad, 1997, for a
historical review). There are a number of possible reasons why a theory might be
sought to explain these associations. Traditionally, however, the theories have most
often been formulated in the context of explaining intellectual development
(Ackerman, 1996). After Cattell first conceived of the concepts of fluid and
crystallized intelligence (Cattell, 1943), the question emerged of how basic raw
ability (Gf) developed into acquired knowledge (Gc). Cattell proposed the
Investment Theory, which specified that Gc was the result of time invested, and of
interest levels for specific areas of knowledge (Cattell, 1987). Vernon (1961) also
theorized that industriousness and general academic interest both contributed
positively to “educational ability”, but early studies were hindered by small sample
sizes and a lack of broad measures for personality and interests.
As models in the three domains improved over time and measures become
more standardized, it became increasingly possible to gather results from diverse
studies. The first true meta-analysis of intelligence-personality associations was
conducted by Ackerman and Heggestad (1997). This study also included a more
qualitative review of the literature on interest-intelligence and interest-personality
associations. These associations were the basis for a new theory of intellectual
development called PPIK theory (intelligence-as-process, personality, interests, and
intelligence-as-knowledge) by Ackerman (1996). The theory shared several
similarities with investment theory, including maintenance of the distinction between
raw or fluid ability (intelligence-as-process) and crystallized intelligence
(intelligence-as-knowledge). Crucially, however, the theory specified that the
overlaps among intelligence, personality and interests took the form of four “trait
complexes”, which were defined as being similar to Snow’s concept of an aptitude
complex in the learning domain (Snow, 1989). Snow proposed there were
combinations of level of traits, such as cognitive abilities, personality traits and
9
motivational traits that statistically interacted to produce better or worse outcomes in
learning situations. Ackerman and Heggestad (1997) extended this concept to
acquisition of academic knowledge more generally. Moreover, they claimed that this
type of specialized knowledge is important to future occupation, a point addressed
below.
A different approach to the integration of individual difference across the
three big domains is to use occupational interests as an underlying framework
(Anthoney & Armstrong, 2010; Armstrong et al., 2008; Armstrong et al., 2011).
Armstrong and colleagues have proposed that the RIASEC model should be used
because of its focus on work environments. They argued that educational and work
environments are crucial to understanding the links among interests, personality and
cognitive abilities, because these environments create demands for these traits, thus
providing key contexts for them to become related (Armstrong et al., 2008). From
the opposite perspective, it is thought that the demands of different occupational
environments “pull” individuals towards them who have traits that would allow them
to meet those demands. Thus, educational and work environments are thought to
both have mutually reinforcing relations with traits, because they both select for the
traits and potentially enhance them once individuals are in the environments. These
ideas are not unique to Armstrong and colleagues, but have been proposed by a
number of theorists on the development of occupational interests (Gottfredson, 2005;
Hogan & Roberts, 2000; Scarr, 1996). One difference, however, is that Armstrong
and colleagues made the specific claim that the RIASEC framework can be used to
understand these relations.
Using a multiple-regression technique called property vector fitting,
Armstrong and colleagues have attempted to fit personality traits and cognitive
abilities onto the Holland hexagon (Anthoney & Armstrong, 2010; Armstrong et al.,
2008). Armstrong et al. (2008) contained three studies; the first used data from
several large studies that related the RIASEC types to personality traits (including
the Big Five) and work styles from the Jackson Vocational Interest Survey (Jackson,
1977). A two dimensional RIASEC circumplex was specified, and property vector
fitting was used to regress the personality and work style scores onto the two
10
dimensions. Of the 51 personality traits and work styles, two-thirds (34) of them had
more than 50% of their variance explained by the dimensions, indicating a good level
of integration into the framework. The distribution of the traits provided support for
both Prediger’s (1982) and Hogan’s (1983) underlying dimensions. In the second
and third studies of Armstrong et al. (2008), cognitive ability was integrated into two
and three dimensions of the RIASEC circumplex, which was also successful for a
majority of the abilities. However, one limitation of these analyses as compared with
Ackerman and Heggestad (1997) was that cognitive ability requirements were rated
for different jobs, but were not derived from intelligence tests. Similarly, in
Anthoney and Armstrong’s (2010) study self-ratings of cognitive abilities were
employed. Thus, there is a need to examine how actual cognitive ability scores fit
into these models. Another limitation of Armstrong and colleagues’ two studies was
that the framework was dependent to a large degree on the RIASEC model, which
likely does not give enough weight to job prestige (Deng et al., 2007).
A third possible integrative approach is to view interests and cognitive
abilities as part of personality, broadly considered. DeYoung (2011) proposed that
intelligence could be located underneath Openness to Experience in the FFM. In a
previous study on the facets of Openness to Experience, DeYoung found that they
were split into two domains: one labelled Openness which consisted of “aesthetically
oriented traits”, and the other called Intellect, which was formed from facets for
intellectual engagement or self-perceived intelligence (DeYoung, Quilty, & Peterson,
2007). General cognitive ability was most strongly correlated with the Intellect
aspect of Openness to Experience, as represented by the Ideas facet on the NEO PI-R
(DeYoung, Peterson, & Higgins, 2005). However, this is a simplified picture
because g has shown many smaller relations with at least one facet for each of the
Big Five (DeYoung, 2011). In addition, narrower cognitive abilities beyond g are
likely to have differential relations with personality traits. For example, Ackerman
and Heggestad (1997) found that Conscientiousness is associated with Conventional
interests in the RIASEC, which are in turn related to Perceptual Speed; thus it would
be predicted that Consciousness is also associated with Perceptual Speed, although
this association has not yet been observed directly.
11
From DeYoung’s (2011) viewpoint, occupational interests fit within
personality at the level of characteristic adaptations. Characteristic adaptation is a
concept taken from the personality theory of McCrae and Costa (2012); it is defined
as an acquired attribute, such as skill or attitude, that arises from the transaction of
the person with the environment. Characteristic adaptations are contrasted with the
basic tendencies that underlie the Big Five, which are thought to be more
biologically-based and resistant to environmental influence. In apparent opposition
to this view, behaviour genetic research has found that vocational interests display
similar heritability coefficients to personality traits (Lykken, Bouchard, McGue, &
Tellegen, 1993). However, Lykken et al. (1993) suggested that much of the
heritability of interests could be explained as an indirect effect of genetic influence
on other attributes, such as physique, personality, and cognitive ability. The
heritability of interests could also be the result of gene-environment interaction and
correlation of more basic traits; for example, if personality affects the initial selection
of learning environments, and the success of individuals in those environments
(Lykken et al., 1993). This hypothesis is echoed in a number of investment theories
(Bouchard, 1997; Gottfredson, 2005; Hogan & Roberts, 2000; Scarr, 1996).
Nonetheless, a major disadvantage of theories explaining occupational interests from
this perspective is that they do not contain the detailed predictions of trait overlap
that are provided in Ackerman’s PPIK theory and Armstrong’s framework.
The three integrative theories of personality, interests and cognitive abilities
can potentially be distinguished by examining how well their models of the overlap
match empirical data. PPIK theory proposes that this overlap is characterized by
four trait complexes that involve groupings of high levels of particular personality
traits, cognitive abilities and interests. The framework of Armstrong and colleagues
instead suggests that personality and cognitive abilities should be mapped as
continuous variables onto the RIASEC model of interests (Armstrong et al., 2011).
The interests are the primary focus because they refer to preferences for education
and work environments, which are theorized to be the contexts in which cognitive
abilities and personality become related to each other and to interests. Finally,
DeYoung (2011) has theorized that cognitive abilities and interests can be integrated
into the FFM model of personality, where cognitive ability is found primarily under
12
Openness to Experience, and interests are characteristic adaptations resulting from
the transaction of personality traits and cognitive abilities with the environment.
The last study of this thesis was focused first on examining the content of
ability-interest trait complexes of PPIK theory. There were two main reasons for
selecting PPIK theory. First, the proposed trait complexes were more parsimonious
and specific than the many possible overlaps between cognitive abilities and interests
in Armstrong’s and DeYoung’s approaches. This made them easier to identify and
potentially falsify. Second, Ackerman and colleagues put forth the hypothesis that
the trait complexes would demonstrate better predictive validity for occupation than
individual scores for the three trait domains. For example, Ackerman and Beier
asked: “is there a synergy among elements within the trait complexes, so that
concentrating on trait complexes is more informative in the career choice context
than individual trait measures?” (2003a, p. 209). This question provided another
prediction of PPIK theory to test.
1.3 Prediction of occupational type
The three integrative approaches presented thus far have been assessed on
their merits as theoretical frameworks, but a key issue is how they could potentially
improve our ability to understand and predict occupational attainment. While the
other theories have not involved as strong a claim for predictive power as PPIK
theory, predicting occupation is a stated goal for most research in this area
(Armstrong et al., 2008). As cognitive ability and personality are both related to and
predict occupation (De Fruyt & Mervielde, 1999; Schmidt & Hunter, 2004), it is
logical to hypothesize that integrative theories could provide superior prediction to
considering each of the domains separately.
PPIK theory has a notable advantage over the frameworks of Armstrong et al.
(2008) and DeYoung (2011), in that it has existed for a longer time, and thus more
research has been done to link the theory to real-world outcomes. Ackerman and
colleagues have found that their trait complexes relate to academic knowledge
(Ackerman & Rolfus, 1999), university course selection (Ackerman, 2000) and
university course performance (Kanfer, Wolf, Kantrowitz, & Ackerman, 2010). The
results were taken as support for PPIK theory because knowledge is hypothesized to
13
act a mediator between trait complexes and occupational attainment (Ackerman,
1996). Nevertheless, these studies only provide indirect evidence for theory because
the predictive validities of either trait complexes or knowledge have not been
examined for attained occupation.
In contrast to the indirect evidence for PPIK theory, there is not yet any
evidence that the approach of Armstrong and colleagues could improve the
prediction of occupation. This would require research relating the framework to
occupational outcomes, possibly comparing its predictive validity to other theories.
Previous research has demonstrated that personality and cognitive abilities can be
mostly effectively integrated into two or three RIASEC dimensions (Anthoney &
Armstrong, 2010; Armstrong et al., 2008). As the RIASEC dimensions are linked
closely with preferences for different educational and occupational environments
(Holland, 1997), the model with these dimensions could be useful in predicting
future occupation, but this remains hypothetical.
DeYoung (2011) has provided a theoretical argument for how cognitive
abilities and interests can be fit into the FFM. However, this account remains very
general and does not specify, for example, which personality traits are involved in
the formation of which occupational interests, or how narrow cognitive abilities fit
into the FFM. Without these details it is not yet possible to use this theory to predict
occupation.
Of the three integrative theories in the literature, PPIK is the most developed.
It has made the most specific predictions for the overlap between cognitive abilities,
personality and interests, and some indirect evidence has been found that trait
complexes relate to occupation. For these reasons, I chose to test this theory in
Project TALENT.
14
Chapter 2: Project TALENT’s design and measures
Project TALENT (PT) was a study approved by the U.S. Department of
Education in 1959 (Flanagan, et al., 1962). It was first conceived by John C.
Flanagan, a professor of psychology at the University of Pittsburgh, who became the
principal investigator. It was designed to be a longitudinal and nationally-
representative study of the human talent of high school students, examining how this
talent could be better identified and promoted. For example, the U.S. Commissioner
of Education stated that the project was “an attempt to determine why so much of the
nation’s human potential is lost and what schools, counselors and parents can do to
reduce the loss” (p. 1, Flanagan, 1962). To this end, a large amount of information
was to be collected about the students (e.g. their aptitudes, interests and social
backgrounds), as well as schools (e.g. their resources and teaching methods). The
following description of the study relies heavily on the first PT report (Flanagan, et
al., 1962). Details of the testing materials are also provided in the Project TALENT
handbook (Wise, McLaughlin, & Steel, 1979). The computerized PT data was
compiled by the American Institutes for Research, a nonprofit social science research
institute founded by Dr. Flanagan. The data is available through the National
Archive of Computerized Data on Aging (NACDA), from which they were obtained
for the current research.2
The advisory panel of Project TALENT and its staff designed the study and
its measures in 1958 and 1959. The schools were selected using a stratified random
sample of public and private high schools across the United States. In all, 1353
schools were eventually sampled (93% of those asked), and approximately 440,000
students, who represented approximately 5% of the total American high school
population. To enable the administration of the tests in each area, 90 regional
coordinators were employed. The regional coordinators were primarily
psychologists who were asked to work with local school administrations and
teachers. Teachers and guidance counselors were trained to administer the Project
TALENT tests, which they gave over two days. The initial testing occurred in
2 The website of the NACDA is http://www.icpsr.umich.edu/icpsrweb/NACDA
15
March of 1960. Follow-up mail surveys were conducted after the students completed
high school, after the intervals of 1, 5 and 11 years. The follow-up surveys asked
about the participants’ personal, educational and career experiences. Most relevant
to the study presented in chapter 5, participants were asked about their current
occupations at those times. In this chapter the data used in the thesis are first
described: the measures of intelligence, personality, occupation interests, and follow-
up occupational status.
2.1 Intelligence tests
The PT aptitude and achievement tests were newly-designed for the study.
Their stated purpose was to “survey a variety of human aptitudes and to obtain scores
which might predict an individual’s ability to develop those aptitudes for vocational
and educational success” (p. 57, Flanagan et al., 1962). One of the main reasons that
new tests were created is that it was felt that pre-existing intelligence tests did not
survey a wide enough variety of aptitudes, partly because the individual subtests
were too long, and the ones in PT should be shorter. In addition, this would make it
certain that none of the students had been previously exposed to the new tests.
A first experimental battery of all the tests was given to a sample of
approximately 6000 high school students, in schools in the Northeast, South and
Midwestern U.S. (Flanagan, et al., 1962, p. 60). Item-level analysis was used to
exclude items that were unreliable, too hard, or too easy. Following this process, the
final 60-test version was developed. The battery was composed of two main
sections: the information tests, and the specific aptitude and achievement tests. For
detailed test descriptions see chapter 3; here their general purpose and design is
outlined.
The information tests were multiple-choice knowledge questions on a very
broad range of topics, including both general knowledge and academic subjects.
There were 36 subtests that ranged from 2 to 24 items in length. There were several
purposes to the information tests. Firstly, it was held that the breadth of a person’s
knowledge was a measure of general intelligence; similar information tests were used
in this way in the Army Alpha and Otis Mental Ability batteries. Second, the more
specific tests had the potential to capture achievement in particular areas, as well as
16
interest and motivation towards those topics, such as physical science, fine art or
sports. Third, there was a vocabulary scale, which was regarded as a measure of
verbal intelligence. In practice, the usefulness of many of the smaller information
subscales was limited because of their narrow topics and poor reliability (Cureton,
1968; Flanagan et al., 1964). Cureton (1968) recommended that tests with less than
nine items be excluded for intelligence research, which eliminated 15 tests. In
addition to this, there were a number of tests that were highly likely to be sex-biased.
For example, tests of Sports and Farming information favoured boys, whereas the
Home Economics tests required knowledge to which girls were more likely to be
exposed. Avoiding unreliability and sex-bias meant that only a maximum of 16 out
of 36 tests (44%) were used in the studies of this thesis.
There were 24 aptitude and achievement tests in the PT battery. The types
and number of aptitude tests were as follows: verbal (3), spatial visualization (2),
reasoning (3), memory (2) and processing speed (4). The achievement tests
included five English tests and three Math tests. The English tests assessed basic
writing and reading skills learned in school, and the Math tests assessed arithmetic,
introductory Math (studied in 9th
grade) and advanced Math (studied in grade 10 or
later). The advanced Math test was left out of all studies here because it was deemed
to be unfair for younger students. Generally, all of the aptitude and achievements
tests were used in assessing cognitive ability in the present research, because most
did not require prior knowledge, and even those that did (e.g. the English tests), only
demanded basic knowledge to which all students should have been exposed.
2.2 Personality tests
There were 150 personality items in PT, which were in a section entitled
“Student Activities Inventory”. Students were asked to respond to statements about
behaviors or characteristics in terms of how well they described “the things I do and
the way I do them”. Reponses were on a five-point Likert scale. The scores on the
items were summed to form 13 scales; however, three of the scales were
experimental and were not electronically recorded. Thus, there were ten personality
scales made up of 108 items, with 7 to 24 items per scale. Item-level data were not
available from the PT dataset, only scale scores.
17
The scales assessed general personality, but were aimed at personality traits
that were important in educational and occupational contexts. For example,
Flanagan et al. (1962) stated that “the TALENT [personality] battery was based on
the hope that it would eventually add to our knowledge of how personality
differences help to account for the differences in accomplishments of equally
talented normal people” (p. 130). Thus, the scales have strong representation of
traits that would be classified under Conscientiousness in the Big Five. However,
Reeve, Meyer and Bonaccio (2006) re-administered the PT items to 219 university
students, and found that the scales spanned all of the Big Five (as assessed by the
NEO-PI-R). In a joint factor-analysis with the NEO, each of the Big Five received at
least one substantial loading (mean r = .70, range = .51 to .81) from the PT scales.
Another difference of the PT personality scales from conventional personality
assessment is that the students were aware that the purpose of the study was to
examine talent, and it was conducted in a school context. Thus, although the
instruction to students was to reflect on their general behavior, the context may have
influenced them to respond in a manner more consistent with how they perceived
themselves within school, or how they wished to portray themselves in a school
setting. This possible confounding factor is explored in greater detail in chapter 4.
2.3 Occupational interest tests
The Interest inventory of PT was composed of 205 items, of which 122 were
occupation titles and 83 were occupation-related activities. Students were asked to
indicate how much they would like to do the occupation or activity. The PT study
designers evaluated pre-existing interest scales, such as Strong’s Vocational Interest
Blank and Kuder’s Preference Record, but decided to construct new scales.
The interest items were compiled into seventeen scales by PT investigators,
based on a priori classification of different occupational areas. Using fifteen
independent raters, Reeve and Hakel (2000) found that all of the PT interest scales
except one (Labour) could be assigned to the RIASEC categories with acceptable
accuracy (inter-rater agreement of 66% or higher). Unlike the personality scales,
however, item-level data were available for the interests. As the original PT interest
18
scales were not created on an empirical basis, in chapter 5 new scales were derived
by factor analysis.
2.4 Occupation at follow-up
The PT follow-up data were collected primarily by the use of mail
questionnaires. The questionnaires were designed to give a broad overview of the
participants’ lives after high school, thus in addition to questions on educational and
occupational experience, they were asked about their marital status, quality of life,
health, and other social variables (Wise et al., 1979).
Of the greatest relevance to the current research were the 11-year follow-ups,
which occurred in 1971 to 1974, when the participants were approximately 28 years
of age. Current occupation was asked in a written response, which was originally
transformed into over 1000 occupation codes. These specific codes were later
reduced to 254 job codes representing specific jobs or job areas such as Airplane
Navigator, Veterinarian or Metal Trades (Wise et al., 1979). The job titles were also
organized into twelve categories according to broad occupational themes. Greater
detail on the occupation categories and the frequencies of participants in each are
provided in chapter 5.
Although efforts were made to contact all PT participants, the participation
rates for each subsequent follow-up decreased. Much of this attrition was due to lack
of the most recent addresses for participants; addresses were lost for approximately
5% of participants for each year, in each grade, compared to baseline. Response
rates also decreased with the time between the baseline testing and follow-ups, which
were longer for the participants who were in lower grades (younger) at baseline. For
the 11-year follow-up, 28.8% of the grade-12 participants returned the
questionnaires, but only 19.6% of the grade-9 sample. To deal with attrition and the
lack of representativeness of the follow-up samples, PT investigators conducted
special interviews with approximately 2500 non-respondents to the mail
questionnaires. The missing participants were found by a variety of methods, such as
searching telephone directories, asking the Department of Motor Vehicles, and
contacting the high school for new addresses. Once participants were located they
were given telephone or in-person interviews (Wise et al., 1979). Sample weights
19
were created in accordance with the sampling ratio of the special sample to original
the 1960 sample (Wise et al., 1979). These sampling weights could then be used to
adjust the follow-up sample to be representative of the baseline sample. The
sampling weights were used in chapter 5 when follow-up occupation was being
investigated.
20
Chapter 3: Comparing the VPR, CHC and Extended Gf-Gc models
3.1 Introduction
Disagreement about the structure of intelligence has a long history in psychology.
Recently, however, some researchers have proposed that a consensus theory has
emerged in the form of the Cattell-Horn-Carroll (CHC) theory of cognitive ability
(Benson, Hulac, & Kranzler, 2010; Flanagan, Ortiz, & Alfonso, 2007; McGrew,
2005, 2009). McGrew (2005), for example, asserted that: “[Carroll’s synthesis] has
finally provided both intelligence scholars and practitioners with the first
empirically-based consensus Rosetta stone from which to organize research and
practice” (p. 171). This view was contradicted, however, by three recent studies in
which an updated version of Vernon’s verbal-perceptual model (1961, 1965) was
found to provide better fit to large intelligence test batteries than the two precursors
of the CHC model: Horn and Cattell’s fluid-crystallized (Gf-Gc) model (Cattell,
1963; Horn & Noll, 1997) and Carroll’s (1993) three-stratum model (Johnson &
Bouchard, 2005a, 2005b; Johnson, Te Nijenhuis, et al., 2007). Nevertheless, these
previous comparison studies relied on an interpretation of Gf-Gc theory which only
included fluid and crystallized ability as second-order factors, whereas the Extended
Gf-Gc theory contains six more such factors (Horn & Blankson, 2005). In addition,
CHC theory has largely supplanted the three-stratum theory, and contains a number
of differences from it (see below, and McGrew, 2009, for details). The current study
was thus aimed at providing an updated test of whether the verbal-perceptual-image
rotation (VPR) model (Johnson and Bouchard, 2005a), the CHC model, or the
Extended Gf-Gc model provides a better description of the structure of intelligence.
Deciding among these models is an important issue for intelligence researchers
because each implies a different underlying theory about the nature of intelligence
and its manifestation. The three models have also each received substantial
empirical support (Carroll, 1993; Horn & Noll, 1997; Hunt, 2011; Johnson &
Bouchard, 2005b). The fluid-crystallized model has arguably been the most
influential theory of intelligence to date in terms of the frequency of its application in
21
research and test development (Carroll, 1993; Horn, 1998; Mackintosh, 2012). As
Kaufman (2012) recently observed “the core concepts of Gc and Gf are still universal
to nearly all IQ tests” (p. 119). It has also been claimed that the CHC model has the
most cumulative factor-analytic evidence supporting it, thanks in large part to
Carroll’s (1993) major synthesis (McGrew, 2009). Although the VPR model has not
been as prominent in the literature as the other two, we argue below that it has a
number of advantages over the Gf-Gc and CHC models.
The main features of the three models are outlined in Table 3.1. These features
are highlighted because they best reflect the theoretical differences among the
models. Although these differences are based upon the most recent versions of the
models, they have their roots in longstanding disagreements about the structure of
intelligence. The CHC and Gf-Gc models are products of the American school of
intelligence research, while the VPR model has its origins in the British school
(Vernon, 1961, Carroll, 1993). In the early twentieth century, the divergent views of
these two schools on the structure of ability were represented by Spearman and
Thurstone. Spearman and his fellow British psychologists such as Burt (but not
Thomson) emphasized the importance of the general factor of intelligence (g) over
group factors in the structure of cognitive ability, whereas American psychologists,
led by Thurstone, supported a model of orthogonal group factors, named primary
mental abilities, with no general factor (Thurstone, 1938). Spearman and his
colleagues argued that Thurstone’s seven to nine primary factors were correlated and
thus could also yield a model with a general factor and smaller group factors
(Eysenck, 1939; Speaman, 1939). Whereas Thurstone rather quickly acknowledged
the presence of higher-order factors in his datasets (Thurstone, 1947, cited in Carroll,
1993), and helped to develop the techniques for higher-order factor analysis, his
reluctance to accept Spearman’s g, and his conception of independent primary
mental abilities had a lasting influence upon American intelligence researchers
(Carroll, 1993). Notably, Cattell and Horn followed Thurstone in not accepting a g
factor in their Gf-Gc model (Horn & Noll, 1997). In contrast, the g factor was
prominent in Vernon’s verbal-perceptual model (1961), and remains so in the VPR
model. Nevertheless, it should be noted that VPR theory is agnostic about whether g
represents a reflective or formative variable. We took the latent factor model
22
approach to the VPR model in the current study, yet the model could be reformulated
to conform to other approaches to the positive manifold (D. J. Bartholomew, Deary,
& Lawn, 2009; Van Der Maas et al., 2006).
Table 3.1
Primary features of the CHC, Extended Gf-Gc and VPR models.
Feature CHC model Extended Gf-
Gc model
VPR model
g factor postulated? yes no yes
Number of second-order
factors
10 (plus 6 more
“tentatively
identified”)
8 3
Second-order factors are
distinguished as content
factors versus raw ability
factors, or by content only.
Content (Gc, Gq,
etc.) versus raw
abilities (Gf, Gsm,
etc.) a
Content (Gc)
versus raw
abilities (Gf,
Gs, etc.) b
Content only
Number and nature of first-
order factors
Pre-specified Pre-specified Left to battery
content
a Gq is quantitative knowledge, Gsm is short-term memory (see McGrew, 2009).
b Gs processing speed (McGrew, 2009).
Although the CHC model does contain a g factor, its second-order factors are
highly similar to those in the Gf-Gc model. This is because the CHC model was
formed by merging Carroll’s (1993) three-stratum model with the Gf-Gc model
(McGrew, 1997, 2005). Carroll himself was also strongly influenced by Gf-Gc
theory, writing that prior to his theory it was “the most well-founded and reasonable
approach to an acceptable theory of the structure of cognitive abilities” (p. 62, 1993).
The original Gf-Gc model had only two second-order factors of fluid and crystallized
ability; however, Cattell and his student Horn eventually added six other second-
order factors. These latter factors resemble Thurstone’s primary abilities: for
instance, quantitative knowledge (Gq), which is similar to the primary ability
numerical facility, visual processing (Gv), which is similar to the primary ability
spatial relations, and processing speed (Gs), which is similar to the primary ability
perceptual speed (Horn & Blankson, 2005; Mackintosh, 2012; McGrew, 2009). Due
to the interdependence of the Gf-Gc and CHC models, analogous factors are present
23
in the CHC model. Thus, the number and nature of the second-order factors in both
models can still be traced back to Thurstone’s primary abilities. The CHC and Gf-
Gc models are distinguished at the second-order level chiefly because the CHC
model has several additional factors, most notably, reading and writing ability (Grw)
and domain-specific knowledge (Gkn); these two factors are instead subsumed by Gc
in the Gf-Gc model (Horn & Blankson, 2005; McGrew, 2009).
P. E. Vernon, who was a contemporary of both Thurstone and Cattell, proposed
the first hierarchical model of intelligence in 1950 (Vernon, 1961). In contrast with
Thurstone’s primary factor model, Vernon’s verbal-perceptual model contained a g
factor and only two broad second-order group factors: the v:ed factor subsumed first-
order factors for verbal, scholastic and numerical ability, and k:m was formed by
loadings from first-order factors of mechanical information, spatial ability, and
perceptual and psychomotor abilities (Vernon, 1961, 1965). Johnson and Bouchard
(2005a) found that the addition of a second-stratum Image Rotation factor
significantly improved the fit of the verbal-perceptual model, and thus they proposed
the Verbal-Perceptual-Image Rotation (VPR) model as an extension of Vernon’s
model. This return towards a more parsimonious model similar to Vernon’s was also
anticipated by researchers such as Undheim (1981) and Gustafsson (1984).
The third feature in Table 3.1 indicates that the broad group factors in the VPR
model are characterized by the subject-matter content of the tests (Johnson &
Bouchard, 2005b). In the CHC and Gf-Gc models there is instead a contrast between
factors which are theorized to involve more basic process abilities (e.g. Gf, Gsm),
and those which are thought to be measures of acquired knowledge (e.g. Gc, Gq)
(McGrew, 2009; Horn & Blankson, 2005). For example, Carroll (1993) stated that:
“the [second-order] domains appear to differ in the relative emphasis they give to
process, content and manner of response” (p. 634). However, as mentioned above,
this distinction between fluid and crystallized factors was not supported in previous
model-comparison studies where the VPR model outperformed Gf-Gc and three-
stratum models (Johnson & Bouchard, 2005a, 2005b; Johnson, Te Nijenhuis, et al.,
2007). In fact, Johnson and Bouchard (2005a) found that even in Cattell’s test
battery designed according to Gf-Gc theory (the Comprehensive Ability Battery;
24
Hakistian & Catell, 1975), the verbal-perceptual distinction was better supported
than the fluid-crystallized one (as assessed by model fit).
The fluid-crystallized division also leads to theoretical problems for the Gf-Gc
and CHC models. In the Gf-Gc model the contrast between ability and knowledge
domains is emphasized to the exclusion of a g factor (Horn & Blankson, 2005);
however, the g factor has been supported in almost all factor-analytic studies where it
was possible to find one (Carroll, 1993; Jensen, 1998) and the Gf and Gc factors
have a correlation as high as .85, supporting an underlying g factor (Johnson &
Bouchard, 2005b). In the CHC model, the presence of both a g factor and a Gf factor
is problematic because of their theoretical similarity: both factors have been
described as involving the ability to reason and profit from experience across many
cognitive domains (Carroll, 1993; Cattell, 1987). For example, Carroll (1993) stated
that: “in the main, I accept Spearman’s concept of g, at least to the extent of
accepting for serious consideration his notions about the basic process measured by
g—the apprehension of experience… and the eduction of relations and correlates” (p.
637), while Cattell wrote of the Gf factor that it was “a single relation perceiving
capacity” that could be invested in any cognitive domain (1987, p. 138, cited in Kan,
Kievet, Dolan, & van der Maas, 2011). This theoretical redundancy has been pointed
out by several authors, and a number of studies have found that g and Gf factors are
statistically indistinguishable (Gustafsson, 1984, 1988, 2002; Kan et al., 2011; Keith,
Fine, Taub, Reynolds, & Kranzler, 2006; Kvist & Gustafsson, 2008; Undheim &
Gustafsson, 1987); however, these models also tend to generate often
unacknowledged out-of-range parameter estimates, such as negative residual
variances. Recently, Kan et al. (2011) also found that, in participants with equal
educational backgrounds, the Gc factor was identical with verbal comprehension,
which is in conflict with the theoretical interpretation of Gc in Cattell’s investment
theory, but consistent with the role of the verbal factor in the VPR model (Johnson &
Bouchard, 2005b).
The last feature which distinguishes the VPR model from the Gf-Gc and CHC
models is the number of and nature of the first-order factors. Along with Spearman
and other early British intelligence researchers, Vernon (1961) criticized American
25
investigators for accepting too many group factors in the lower orders of the
intelligence hierarchy; he argued that this was due to overly lax selection criteria and
because their factor-analytic methods assigned some of the g variance to group
factors. In favor of his more limited set of broad group factors, Vernon noted that
v:ed and k:m emerged in any representative battery of tests, whereas the narrower
(generally, first-order) factors proposed by American psychologists were very
dependent on the particular tests administered and the selectiveness of the sample
(see Appendix in Vernon, 1961). This lack of certainty about narrow factors is
maintained in the VPR model, in that it does not make specific predictions about
which first-order factors should emerge in a given test battery, instead leaving the
characters of the factors to vary according to the specific tests in the battery (Table
3.1). Vernon (1961) also offered a pragmatic argument against naming and
including narrow ability factors in the structure of intelligence; he observed that
often the narrow factors in intelligence test batteries did not add substantial
incremental variance to the prediction of educational or occupational performance,
over and above g and the broad group factors. This objection is not taken into
account by CHC investigators, who aim to include every factor identified in
intelligence research in the CHC theory/model (McGrew, 2009), regardless of
whether they add significant incremental validity over higher-order factors towards
predicting outcomes of interest. Proponents of the Gf-Gc model also maintain that
first-orders factors should be named, and that the factors which should emerge for a
given battery can be pre-specified (Horn & Blankson, 2005).
In spite of claims for its status as the leading intelligence theory (McGrew, 2009),
there are still empirical reasons to doubt whether the CHC model provides an
accurate picture of the overall structure of intelligence. Carroll’s (1993) three-stratum
theory was based on his interpretation of numerous exploratory factor analyses, not
on confirmatory factor analysis, which allows the researcher to investigate and
control many more aspects of the measurement model, and, especially, to pit
competing models against each other empirically. Second, the vast majority of the
datasets re-analyzed by Carroll were not suited to determining the broad higher-order
structure of ability: the CHC model contains at least ten second-order factors, but all
except two of Carroll’s 461 datasets contained three or fewer second-order factors.
26
Although the broad CHC factors have been supported in a number of more recent
studies (see McGrew, 2005, for summary), much of this research has been performed
on test batteries designed within the CHC framework or its precursors, and
competing models have not been compared with it. These criticisms also apply to
the factor-analytic evidence supporting the Extended Gf-Gc model (Horn &
Blankson, 2005).
In order to establish whether the CHC, Extended Gf-Gc, or VPR model is the
best-supported, further confirmatory studies are needed which compare their
predictions in test batteries that were not constructed according to any particular
theory of intelligence. As mentioned above, previous studies have not examined the
most recent versions of these models, thus this was the main purpose for the current
study.
3.1.1 Previous factor-analytic research on Project TALENT
The current study was undertaken with data from Project TALENT, which was
a longitudinal study on American high school students that was designed to
investigate their aptitudes, interests, and backgrounds, and the influences of these
variables on educational and occupational outcomes (PT; Flanagan et al., 1962).
During Project TALENT, 60 aptitude and achievement tests were given to a very
large and nationally-representative sample of the U.S. student population (see
Methods below for more details). As detailed below, three PT datasets were also
analyzed by Carroll (1993), which provided a basis for the development of the CHC
model in our study. The data are thus of particular relevance to the question of
which of the three models provides the best fit. In order to provide context for the
factor analysis of this test battery in this study we first review notable previous
analyses of these data.
The first report which contained an analysis of the aptitude and achievement
tests in PT was written by the research group who designed the study (Flanagan et
al., 1964). Instead of performing factor analysis, however, Flanagan et al. (1964)
examined correlation matrices and uniqueness coefficients of the tests. Using this
method they tentatively identified and labeled seven common factors: general verbal
27
ability, reasoning, rote memorization, spatial visualization, visual perception, speed
of response, and information in the mechanical-electric-electronic domain.
Lohnes (1966) ran principal components analysis on a combined sample of
grades 9 and 12 participants from Project TALENT. He used all 60 test scores in the
battery and extracted eleven factors, excluding factors for grade and sex. However,
four of these factors were highly specific (such as information on etiquette, or
hunting and fishing); subsequent investigators have typically excluded a number of
these information tests because of their highly specific nature, but also because they
contained a small number of items and many had low reliability coefficients (see
Flanagan et al., 1964).
Shaycroft (1967) examined the changes in 47 PT test scores from grade 9 to
grade 12, and performed principal-axes factor analysis on the tests. She retained
seven broad factors, as Lohnes (1966) did. However, the most extensive factor
analysis of PT tests in this period was performed by Cureton (1968), who provided
detailed comparisons of his factors to those in Lohnes (1966) and Shaycroft (1967).
Cureton’s (1968) sample consisted of 543 students from Project TALENT who
also completed three other intelligence test batteries. He performed three different
factor analyses: on all the tests combined, the non-Talent tests only, and the PT tests
only. For the PT test analysis Cureton (1968) excluded all the information tests with
less than nine items, and ran principal-axes factor analysis with oblique rotation. He
accepted seven factors, and although these differed slightly by sex, each model
included a factor which combined the English and Math tests, a verbal-information
factor, a clerical-perceptual factor, as well as factors for spatial reasoning,
mechanical/outdoor knowledge, math and memory. These factors were generally
consistent with those in Lohnes (1966) and Shaycroft (1967), despite different
factoring methods and selection of tests in the three studies. Importantly, Cureton
(1968) also observed that the mechanical factor had a tendency to combine with the
spatial factor, and that the verbal-information factor was closely related to the
English and Math factor; thus Cureton (1968) observed that “though second-order
and hierarchical analysis was not used, the results are in striking accord with the
theory of cognitive abilities outlined by Vernon” (p. 71).
28
When Carroll (1993) re-analyzed the data from Project Talent, he revisited the
analyses of Flanagan et al. (1964), Shaycroft (1967) and Cureton (1968), which is a
reflection of the lack of intervening factorial research after these seminal studies.3
Carroll (1993) accepted seven first-order factors in his re-analysis of the grade 9 data
from Shaycroft (1967), and his factors were very similar to those found by Shaycroft
and Cureton, except that he did not find the Math and English tests to combine to
form a factor. Despite the fact that these seven first-order factors were underneath
four different second-order broad factors according to the three-stratum model,
Carroll (1993) obtained only one second-order factor for the male data, which he
classified as Gc (see dataset codename SHAY01; Carroll, 1993). In the female data
(SHAY02), Carroll extracted three second-order factors: 2H, 2V (broad
visualization) and a technical knowledge factor; he also found a third-order g factor.
Carroll’s (1993) analysis of Flanagan et al. (1964) was based on a correlation matrix
which did not include the information tests (FLAN01). He extracted five first-order
factors from these tests: verbal ability (V), math knowledge (KM), English language
usage (EU), visualization (VZ) and perceptual speed (P). According to the three-
stratum model, these factors should have loaded onto three separate higher-order
factors, but Carroll (1993) only obtained one factor, which he characterized as 2H (a
combination of fluid and crystallized intelligence). Together, the re-analyses by
Carroll suggest that the higher-order structure of the PT tests is more parsimonious
than implied by three-stratum theory, and thus potentially Gf-Gc and CHC theory as
well.
Three more recent studies using PT data took Carroll’s (1993) factor solutions
as a starting point (Reeve, 2004; Reeve & Heggestad, 2004; Reeve et al., 2006).
Although Reeve and colleagues found acceptable fit for their confirmatory factor
analysis (CFA) measurement models, these studies were not primarily aimed at
investigating the structure of the PT test battery, and contained no exploratory factor
analysis (EFA) to determine the number of first-order factors for the selected tests,
nor higher-order factor analysis.
3 Caroll’s (1993) re-analysis of Cureton (1968) was based on the dataset with the PT tests combined
with three additional test batteries, and hence is not as relevant to the current review as Carroll’s two
other re-analyses.
29
In the present study we first performed EFA in order to establish the first-order
structure of the PT tests. This provided an objective basis upon which to perform
higher-order CFA to test the relative fits of the CHC, Extended Gf-Gc and VPR
models.
3.2 Methods
3.2.1 Sample
The participants in Project TALENT (PT) were drawn from a stratified
random sample of all public and private high schools in the United States in 1960
(Flanagan et al., 1962). The full obtained sample consisted of 376,213 students, with
approximately 100, 000 students in each grade from 9 through 12. Of the full
sample, 50.13% was female. The age range was from a mean of 14.4 in grade 9 (SD
= .78) to 17.3 in grade 12 (SD = .67). The full individual age range was from 8 to
21.
3.2.2 Measures
Short descriptions and reliabilities of each cognitive ability test used in the
current study are presented in Table 3.2. In addition to aptitude tests, the designers
of PT included a large number of multiple-choice information tests because they
sought to use these to predict future educational and vocational success in a wide
variety of areas (Flanagan et al., 1962). The information tests were based partly on
knowledge acquired from formal education, but were also designed to assess self-
motivated learning outside the classroom (Flanagan et al., 1962). The information
tests were also designed to be non-redundant with the achievement tests; thus the
math information test contained factual items on mathematical concepts, but did not
require problem solving as did the arithmetic and math achievement tests (Flanagan
et al., 1962).
Following previous analysts such as Cureton (1968), we excluded tests with
less than eight items because of their low reliabilities and tendency to form highly
specific factors. An effort was also made to exclude information tests that were
likely to be sex-biased due to unequal learning opportunities for boys and girls, such
as the Sports, Farming and Home Economics tests. Nonetheless, the Aeronautics and
30
Space, Electricity and Electronics and Mechanics tests were retained because of their
importance in distinguishing the VPR from the CHC and Gf-Gc models; the
perceptual factor in the VPR model is formed from a combination of spatial and
mechanical-knowledge factors, whereas these factors load onto separate second-
stratum factors in the CHC and Gf-Gc models (Johnson & Bouchard, 2005b; Horn &
Blankson, 2005; McGrew, 2009). The advanced mathematics test (R333) 4
was
excluded because it included material that was not taught until grades 10 through 12
(Wise et al., 1979); thus it was deemed to be an unfair test for grade 9 students.
The remaining PT battery still contained 16 information tests that we
considered were possibly less relevant to cognitive ability than the aptitude tests.
Flanagan et al. (1962) defended the information tests as indicators of general
intelligence based on their inclusion in classic intelligence test batteries such as the
Army Alpha test and the Otis Mental Ability Tests, but they also noted that
information tests are measures of interest and past achievement in specific areas
(such as Biology, Physics, Literature, etc.). To defuse this question about the
appropriateness of including the information tests, we fit our models to two
selections of tests (hereafter termed the broad and narrow selections). The broad
selection included the information tests, and consisted of 37 tests in total. The
narrow selection excluded the information tests except for Vocabulary, and consisted
of 22 tests.
4 Variable ID numbers that were assigned by Project TALENT are occasionally referenced, in order to
clarify which tests are being discussed.
31
Table 3.2
Project Talent test names, short descriptions, and reliabilities for males/females.
Test Name Description Items Reliability
Vocabulary General vocabulary questions. 21 .71/.71
Literature Items on a broad selection of literary works. 24 .72/.70
Music Musical information (not requiring formal training in music). 13 .67/.67
Social Studies Items on facts and concepts from history, economics, civics,
geography and current affairs.
24 .83/.79
Mathematics Items on mathematical information and concepts. 23 .81/.78
Physical Science Items about chemistry, physics, astronomy, and other
physical sciences, not necessarily acquired through formal
education.
18 .77/.72
Biological Science Questions about botany, zoology and microbiology. 11 .57/.51
Aeronautics and Space Items on flying technique, navigation, jet planes, and space
exploration
10 .63/.34
Electricity and
Electronics
Items on the construction and maintenance of electrical or
electronic equipment.
20 .76/.43
Mechanics Information on automobiles, common machines, etc. 19 .66/.48
Art General knowledge about art, artists and art works. 12 ..64/.65 a
Law General knowledge items that could be acquired through
books or news reports on legal affairs.
9 .51/.43 a
Health Items on practical health maintenance, nutrition and common
health care techniques.
9 .58/.55 a
Bible General knowledge about the characters and teachings of the
Bible.
15 .74/.73 a
Theatre and Ballet General terms from theatre and ballet. 8 .55/.59 a
Miscellaneous Miscellaneous knowledge questions. 10 .48/.42a
Memory for sentences Recalling a missing word from a memorized sentence. 16 .62/.63b
Memory for words Recalling an English word that corresponds to a word in a
(fictional) foreign language.
24 .80/.83b
Disguised words The ability to use phonetic sound to puzzle out which familiar
English word a nonsense word corresponds to.
30 .86/.87c
Spelling Items testing the ability to spell, and not the size of
vocabulary.
16 .60/.56
Capitalization Items requiring the correct capitalization of words in a
sentence.
33 .85/.83
Punctuation Items on the appropriate use of punctuation. 27 .72/.73
English usage Knowledge of preferred phrasing in English. 25 .56/.49
Effective expression Items testing the ability to recognize whether an idea has been
expressed clearly, concisely and smoothly.
12 .63/.52
32
Word functions in
sentences
A test of sensitivity to grammatical structure. The test taker
must find the word that performs the same grammatical
function as a word in another sentence.
24 .81/.84
Reading
comprehension
Multiple-choice items on the comprehension of a written
passage.
48 .86/.84c
Creativity Verbal items requiring ingenious solutions to practical
problems.
20 .73/.68
Mechanical reasoning Items of the ability visualize the operation of physical force,
such as the effect of gravitation, gears, pulleys, levers, etc.
20 .76/.64
Visualization in 2
dimensions
Items requiring mental rotation of shapes in two dimensions. 24 .81/.80c
Visualization in 3
dimensions
Items on the ability to visualize a how a two-dimensional
figure would look after it were folded into a three-
dimensional one.
16 .70/.59
Abstract reasoning A non-verbal test on the ability to identify the logical
progression of elements in a complex pattern (similar to
Raven’s progressive matrices).
15 .66/.65
Math 1 Arithmetic
reasoning
A test of the ability to reason in the manner required to solve
arithmetic problems, with only very simple computation.
16 .73/.71
Math 2 Introductory
High School
mathematics
A test of mathematics taught up to an including the 9th
grade,
including items on algebra, fractions, simple geometry, etc.
24 .78/.73
Arithmetic
computation
A test of the speed and accuracy of addition, subtraction,
multiplication and division.
72 Not avail.
Table reading A test on the speed and accuracy of obtaining information
from a table.
72 Not avail.
Clerical checking A test on the speed and accuracy of checking whether two
pairs of names are identical.
74 Not avail.
Object inspection A test on the speed and accuracy of spotting small differences
between objects when comparing them visually.
40 Not avail.
Note: Descriptions adapted from Wise et al. (1979). Reliability estimates taken from Flanagan et al. (1964,
Table 2-5), and are based on the Kuder-Richardson Formula 21 (Kuder & Richardson, 1937) unless
otherwise noted. The reliabilities are lower-bound estimates, and are based on the mean for all grades
combined. a
Estimate based on Kuder-Richardson Formula 20. b
Estimate may be an overestimate due to lack of experimental independence of items (Flanagan et al.,
1964). c
Split-half reliability estimate.
3.2.3 Data preparation
Among the PT tests was a screening test consisting of basic knowledge
questions that were taught in elementary school; it was designed to identify students
33
who were functionally illiterate, mentally retarded, or who displayed an apathetic
attitude to the tests (Wise et al., 1979). A response credibility index was available
for all participants according to their scores on the screening test, taking into account
whether the score could be explained by illiteracy (a low score on the reading
comprehension test—R250), mental slowness (a low score on the clerical checking
test—A430), problems with clerical inaccuracy (a low percentage correct on the
clerical checking test—P430), or some combination of these. We removed cases
who scored below the threshold for the screening test, except those cases for which
no explanation was provided by their scores on the other three tests. Also, in order
not to restrict the range of cognitive ability, participants who failed the screening test
ostensibly due to mental slowness were left in the sample. Students with missing
scores on the screening test were also retained.
Prior to the analysis, data were screened for normality and outliers. Three
tests were found to have problematic violations of normality in each grade:
Capitalization and English Usage were found to be negatively skewed, and Table
Reading was positively skewed (all three also displayed leptokurtic distributions).
To deal with these violations, a logarithmic transformation was applied to
Capitalization, a square-root transformation to English Usage, and logarithmic and
cosine transformations were applied to Table Reading. The scores for Capitalization
and English usage were reflected prior to transformation, and re-reflected afterwards
in order to keep the original direction of the scores. The scores for Table Reading
were also re-reflected after transformation because the cosine transformation
reflected them. Following transformation, within each grade the highest remaining
skewness was for Clerical Checking (z = 0.70-1.08) and the highest remaining
kurtosis was for Table Reading (z = 1.74-2.63).
After transformation, no extreme univariate outliers remained given the large
sample size. In order to control for potential multivariate outliers the Mahalanobis
distance and Cook’s distance were obtained for complete cases (separately in the
broad and narrow selections). These statistics were obtained by regressing the PT
student ID number (a random variable) onto the test scores. Cases that had a
Mahalanobis distance with a p < .001 (χ2(37) = 69.35, for the broad selection, χ
2(22)
34
= 48.27, for the narrow selection), and a Cook’s distance of greater than 4/N were
removed from the sample (critical values suggested by Tabachnick & Fidell, 2007).
Following data screening, total sample size was reduced to 366,857 in the broad
selection, and 366,695 in the narrow selection (2.49 - 2.54 % of the sample
removed).
We handled missing data by using multiple imputation with five datasets for
the exploratory analyses and direct maximum likelihood estimation for the
confirmatory factor analyses. These missing-data methods yield the same results
(Brown, 2006), and require the assumption that the data be missing at random
(MAR). For this assumption to have been violated, students would have had
selectively to avoid particular tests specifically due to awareness of lower ability in
those areas. Given that 2.27-3.23% of all test scores were missing across each grade,
usually in relatively large ‘clumps’ for individuals, this is unlikely to have occurred
enough to affect the results. A comparison of means and correlations in the full
dataset to those with listwise deletion also showed only very small differences,
indicating that the pattern of results would be the same basically no matter how
missing data were treated.
3.2.4 Analysis method
Despite the existence of previous such analyses in PT, we used exploratory
factor analysis to estimate the factor structure given our particular selections of tests
and data screening methods (for example, our treatment of outliers and missing data).
Consistent with previous analyses of the PT data, factor analyses were performed
separately for each combination of grade and sex (Carroll, 1993; Reeve et al., 2006).
This was done to retain a number of replication samples and to identify possible
differences in the factor structures across the sexes and grades. In order to determine
the numbers of factors to extract in the exploratory analyses, parallel analysis and
Velicer’s minimum average partial (MAP) test were obtained for each sample in
SPSS (see O'Connor, 2000, for syntaxes); the Kaiser criterion (the number of
eigenvalues > 1) was also examined. Most important, however, was a consideration
of the interpretability of the factors and whether each factor had at least two tests
whose highest loadings were there (a criterion suggested by Carroll, 1993). All
35
analyses were performed with maximum likelihood estimation, and Promax rotation
(Kappa set to 4) was used for all the exploratory analyses.
For the CFA results, we report three conventional fit indices: the Root Mean
Square Error of Approximation (RMSEA), Comparative Fit Index (CFI) and
Standardized Root Mean Square Residual (SRMR) (Brown, 2006). An information
criterion index, the saBIC (sample size-adjusted BIC), was selected in order to be
able to directly compare our non-nested models. The saBIC was chosen because the
unadjusted BIC imposes a high penalty for additional parameters based upon sample
size (Kenny, 2011); the saBIC has also been found to perform well in model
selection (Sclove, 1987; Yang, 2006). Although not shown, AIC (Akaike
Information Criterion) statistics gave rise to the same conclusions as the saBIC.
3.3 Results
3.3.1 Exploratory factor analysis
3.3.1.1 Broad selection
For the males in the broad selection, parallel analysis and the MAP test each
indicated that there were four factors, while the Kaiser criterion suggested five
factors (except in grade 9 males, where it suggested four). However, when a sixth
factor was extracted it was clearly interpretable as the math knowledge factor that
was identified by previous researchers (Cureton, 1968; Carroll, 1993); the factor had
the highest loadings for Math Information and the second part of the Math test. In
contrast, a seventh factor was not clearly interpretable and contained no highest
loading from any test. Thus, we retained six factors in each male sample.
In the female samples, parallel analysis suggested there were only three
factors, but the MAP test indicated four factors in grades 9 and 10, and five in grades
11 and 12. The Kaiser criterion suggested five factors throughout. Nonetheless, the
sixth factor was identified as the Math factor in the same manner as in the male
samples. In grades 9 and 10 the seventh factors were singlets with the highest
loadings for Disguised Words and Memory for Sentences, respectively. In grades 11
and 12, the seventh factors had two highest loadings (Memory for Sentences and
Bible in grade 11, and Biological Sciences and Bible in grade 12), but these factors,
36
unlike the others, varied in each sample and were difficult to interpret; thus we
retained six factors in each female sample as well.
Table 3.3 presents the factor pattern matrices for grade 10 males and females,
with loadings under .15 suppressed (.15 was used as a cutoff to determine whether a
factor loading was substantively meaningful). Table 3.5 contains the factor
correlation matrices. The six factors were labelled in a manner similar to Cureton
(1968): the Information factor was formed largely by loadings from the information
tests. The English/Math factor had loadings from the English and Math achievement
tests. The Spatial/Reasoning factor was formed by tests with visual-spatial content,
but also tests that involved a reasoning component (such as Creativity, Math 1). The
Mechanical/Science factor had loadings from tests requiring knowledge in
mechanics, electronics and science subjects. The Speed factor was formed by
loadings from all the speeded tests in the battery. Finally, the Math factor had its
highest loadings from tests requiring math knowledge (as opposed to tests of
computation skills, which loaded more highly upon the English/Math factor).
37
Table 3.3
Factor pattern matrices for grade 10 males/females in the broad selection of PT tests.
Test Name Factor
Inform-
ation
English/
Math
Spatial/
Reasoning
Mechanical/
Science Speed Math
Vocabulary .55/.56 –/.16 .28/.20
Literature .93/.84
Music .75/.74
Social Studies .73/.62
Mathematics .26/.21 .54/.64
Physical Science .35/– .40/.52 .24/.16
Biological Science .41/.23 .31/.40
Aeronautics and Space .48/.33 .37/.33
Electronics .74/.68
Mechanics –/.18 .75/.55
Art .89/.90
Law .60/.55
Health .53/.45 .17/.29 .16/.17
Bible .69/.49
Theatre and Ballet .78/.88
Miscellaneous .65/.57
Memory for sentences .29/.36
Memory for words .17/.15 .34/.38
Disguised words .33/.37 .42/.43 .24/.25
Spelling .72/.81
Capitalization .63/.71
Punctuation .72/.76
English usage .62/.62
Effective expression –/.15 .57/.54
Word functions in sent. .42/.40 .16/.15 .31/.33
Reading comprehension .58/.54 .33/.33 .15/.15
Creativity .28/.29 .19/– .29/.32 .15/–
Mechanical reasoning .62/.69 .36/.15
Visualization in 2D .58/.59 .20/.18
Visualization in 3D .78/.77
Abstract reasoning .17/.25 .58/.57
Math 1 .38/.35 .18/.23 .27/.24
Math 2 .32/.26 .59/.62
Arithmetic comp. .52/.63 .30/.26 .26/.17
Table reading –/.17 .66/.64
Clerical checking .72/.72
Object inspection .30/.27 .62/.63
38
As seen in Table 3.3, the general pattern of loadings was highly similar across
the sexes, with some minor exceptions. The loadings for physical science, biology
and aeronautics tests on the Information factor were lower in females, which may be
attributable to their lower reliabilities in females (see Table 3.2). However, the
loadings of these tests on the Mechanical/Science factor were similar in each sex,
suggesting that they functioned equally well as tests of specific mechanical/science
knowledge in females as in males, but that they were better tests of general
knowledge for males than females. Another interesting sex difference was that the
test of mechanical knowledge loaded at .18 on the Spatial/Reasoning factor in
females, but below .15 in males. This suggests that mechanical knowledge was tied
more closely to spatial ability in females.
Differences in salient loadings across the eight samples were later used for
the construction of the confirmatory factor models. The differences were as follows:
Relative to grade 10 males, in grade 9 males the Vocabulary test had a
loading on English/Math, Reading Comprehension did not have a loading on
Spatial/Reasoning, Creativity did not load substantively (had a loading below .15)
onto the Mechanical/Science factor, and Punctuation had a loading on the Math
factor. In grade 11 males, Memory for Words did not load substantively on the
Information factor. In grade 12 males, the Vocabulary test had a loading on
English/Math and the Electricity and Electronics test had a loading on the Math
factor.
Relative to grade 10 males, in grade 10 females Effective Expression loaded
on the Information factor, while Physical Science Information did not5, Vocabulary
and Table Reading loaded on English/Math, while Creativity did not, and
Mechanical Information loaded on Spatial/Reasoning.
In grade 9 females, relative to grade 10 females, Memory for Words and
Effective Expression did not load on Information, Bible had a loading on
5 This was a logical finding given that most students were probably not exposed to formal Physics
classes until higher grade levels. However, unlike in the female samples, in grade 9 and 10 males
Physical Science did load on Information, perhaps because boys were more likely to be exposed to
Physics knowledge outside of school.
39
English/Math, Word Functions in Sentences and Reading Comprehension did not
load on Spatial/Reasoning, Mechanical reasoning did not load on the
Mechanical/Science factor, and Social Science information loaded on the Math factor
but Physical Science did not.
In grade 11 females, relative to grade 10 females, Physical Science and
English Usage loaded on Information, Table Reading did not load on English/Math,
and Health information did not have a loading on Mechanical/Science.
Finally, in grade 12 females, relative to grade 10 females, Physical Science
and English Usage loaded on Information, while Memory for Words did not,
Vocabulary and Table Reading did not load on English/Math, and Vocabulary and
Health information did not load on Mechanical/Science.
3.3.1.2 Narrow selection
In males, parallel analysis and the MAP test indicated three factors. The
Kaiser criterion suggested three factors in grades 9 and 10, and four factors in grades
11 and 12. In females, all the criteria suggested three factors. Nonetheless, five
factors were retained for both sexes because the fourth and fifth factors corresponded
to factors in the broad selection, and contained the highest loadings from at least two
tests.
The sixth factor was not retained for multiple reasons. In each male sample
the sixth factor was a singlet with Memory for Sentences, and thus was clearly not
interpretable. In grade 9 females, there were no highest loadings on the sixth factor;
in grade 10 there was a doublet with Memory for Words and Memory for Sentences,
and in grade 11 and 12 females there was once again a singlet with Memory for
Sentences. Although there was some evidence for a Memory factor in females, we
considered that the relation between the memory tests would be best handled with
correlated error variances instead of a factor, because of the similarity in the format
of the tests (memorization of words or sentences, followed by multiple-choice items
testing recall).
Table 3.4 displays the factor pattern matrices for grade 10 males and females,
and Table 3.5 the factor correlations. Four of the factors were highly similar to those
40
in the broad selection and were labelled the same: English/Math, Spatial/Reasoning,
Speed and Math. Due to the removal of the science and mechanical information
tests, there was no longer a factor for them. The fifth factor had the highest loadings
for the Vocabulary and Creativity tests and was labelled the Verbal factor because all
its loadings came from tests with verbal subject-matter content.
Table 3.4 Factor pattern matrices for grade 10 males/females in the narrow selection of PT tests.
Test Name Factor
English/Math
Spatial/ Reasoning Speed Math Verbal
Vocabulary .36/.43 .50/.48
Memory for sentences .18/.23
Memory for words .38/.38 –/.15
Disguised words .58/.60 .26/.27 .27/.24
Spelling .79/.81
Capitalization .69/.71
Punctuation .81/.77
English usage .71/.72
Effective expression .60/.59
Word functions in sent. .44/.40 .29/.30
Reading comprehension .51/.52 .45/.44
Creativity .22/.20 .24/.24 .40/.37
Mechanical reasoning .72/.66 .18/–
Visualization in 2D .60/.60 .20/.17
Visualization in 3D .78/.76
Abstract reasoning .21/.28 .54/.53
Math 1 .21/.24 –/.16 .43/.38 .15/.16
Math 2 .17/.23 .68/.58
Arithmetic comp. .28/.38 .30/.25 .44/.39
Table reading .66/.66
Clerical checking .73/.73
Object inspection .28/.26 .62/.63
Relative to grade 10 males, the only factor loading differences in males were
that Math 1 loaded on Spatial/Reasoning in grades 11 and 12.
Relative to grade 10 males, in grade 10 females, Math 1 had a loading on
Spatial/Reasoning, Memory for words had a loading on Verbal, and Mechanical
reasoning did not load on the Verbal factor.
41
In grade 9 females, relative to grade 10 females, Memory for Words had a
loading on the Math factor (allowed only in the VPR model), and Memory for Words
and Math 1 did not load on Verbal.
In grade 11 and grade 12 females, relative to grade 10 females, Math 2 did
not load on English/Math, and Memory for Words and Math 1 did not load on
Verbal.
Table 3.5 Factor correlation matrices for grade 10 males (below diagonal) and females (above diagonal)
in the broad and narrow selections of PT tests.
Info.
English/
Math Spatial/ Reasoning Math Speed
Mech./ Science
Information – .783 .659 .641 .093 .722
English/Math .760 – .676 .688 .196 .570
Spatial/Reas. .593 .593 – .571 .186 .602
Math .612 .607 .500 – .064 .614
Speed .099 .201 .140 .086 – .008
Mech./Science .692 .482 .642 .504 -.037 –
Narrow Selection
Verbal – .582 .595 .513 .070
English/Math .590 – .626 .694 .267
Spatial/Reas. .597 .561 – .593 .294
Math .494 .737 .577 – .175
Speed .068 .211 .163 .176 –
3.3.2 Confirmatory factor analyses
3.3.2.1 Broad selection
Based upon the results of the exploratory analysis, we developed
confirmatory models for the VPR, CHC and Gf-Gc models (see Figures 3.1-3.3).
Table 3.6 displays the loadings of the tests on the first-order factors for the grade 10
male sample. Differences in factor loadings compared to this sample were noted
above in section 3.1.1. Although the number of factor loadings varies across the
models, the number of input variables was the same for each. Supplemental tables
A1 and A2 provide the numerical first-order loadings for the grade 10 males and
females for all models (see Appendix A).
42
In order to represent the three models under consideration accurately, factor
loadings were only included if they were consistent with the theories behind the three
models, taking into account whether test content could account for any indicated
cross-loadings. For the Gf-Gc and CHC models, for example, there was no
theoretical or content rationale for the math achievement and arithmetic tests to load
onto a factor otherwise dominated by English tests (these factors were characterized
as English achievement in the CHC model (following Carroll, 1993)6, and Verbal
comprehension in the Gf-Gc model). A hybrid first-order English/Math factor is also
not specified in the models, and the first-order factors for quantitative ability and
English achievement are theorized to load onto separate second-order factors in both
(Horn & Blankson, 2005; McGrew, 2005). In contrast, a factor combining English
and Math tests is theoretically plausible within the VPR model because factors
formed by these tests are in the same domain, underneath the broad Verbal (formerly
verbal-educational) factor (Johnson & Bouchard, 2005b), and because the VPR
model does not pre-specify factor content. Additionally, three verbal tests were
found to load onto the Math factor (Word Functions in Sentences, Memory for
Words, and Punctuation). These cross-loadings were included in the VPR model
because of the theoretically-based connection between the Math and English factors,
but were not allowed in the CHC and Gf-Gc models.
6 See dataset SHAY01 in Carroll (1993). This factor was also characterized as English language
usage in the other PT datasets, but the label of English achievement seemed more appropriate to us.
43
Table 3.6
First-order loadings for the CHC, Extended Gf-Gc and VPR models in the broad selection (grade 10 males).
Test Primary loadings Secondary loading(s)
CHCa
Gf-Gcb
VPR CHC Gf-Gc VPR
Vocabulary K0 Vi Information K1 Science Mechanical/Science
Literature K0 Vi Information
Music K0 Vi Information
Social Studies K0 Vi Information
Mathematics KM Gq Math K0 Vi Information
Physical Science K1 Science Mechanical/Science K0, KM Vi, Gq Information, Math
Biological Science K0 Vi Information K1 Science Mechanical/Science
Aeronautics and Space K0 Vi Information K1 Science Mechanical/Science
Electronics K1 Science Mechanical/Science
Mechanics K1 Science Mechanical/Science
Art K0 Vi Information
Law K0 Vi Information
Health K0 Vi Information A6, K1 V, Science English/Math,
Mechanical/Science
Bible K0 Vi Information
Theatre and Ballet K0 Vi Information
Miscellaneous K0 Vi Information
Memory for sentences A6 V English/Math
Memory for words A6 V English/Math K0 Vi Information
Disguised words A6 V English/Math K0, P Vi, P Information, Speed
Spelling A6 V English/Math
Capitalization A6 V English/Math
Punctuation A6 V English/Math
English usage A6 V English/Math
Effective expression A6 V English/Math
Word functions in sent. A6 V English/Math Vz Visualization Math,
Spatial/Reasoning
Reading comprehension K0 Vi Information A6, Vz V, Visualization English/Math,
Spatial/Reasoning
Creativity Vz Visualization Spatial/Reasoning K0, A6, K1 Vi, V, Science Information,
English/Math,
44
Mechanical/Science
Mechanical reasoning Vz Visualization Spatial/Reasoning K1 Science Mechanical/Science
Visualization in 2D Vz Visualization Spatial/Reasoning P P Speed
Visualization in 3D Vz Visualization Spatial/Reasoning
Abstract reasoning Vz Visualization Spatial/Reasoning A6 V English/Math
Math 1 KM Gq English/Math Vz Visualization Math,
Spatial/Reasoning
Math 2 KM Gq Math A6 V English/Math
Arithmetic comp. KM Gq English/Math P P Speed, Math
Table reading P P Speed
Clerical checking P P Speed
Object inspection P P Speed Vz Visualization Spatial/Reasoning a K0 = general verbal information, K1 = science knowledge, A6 = English achievement, KM = math achievement, P = perceptual speed.
b Vi = general
information, V = verbal comprehension, Gq = quantitative knowledge.
45
Fig. 3.1. Measurement model of the VPR model with factor loadings from the grade 10 male sample.
46
Fig. 3.2. Measurement model of the Extended Fluid-Crystallized model with factor
loadings from the grade 10 male sample.
Fig. 3.3. Measurement model of the Cattell-Horn-Carroll model with factor loadings
from the grade 10 male sample.
47
According to the CHC framework, the six first-order factors that we obtained each
belonged underneath separate second-order factors (McGrew, 2009). Unfortunately,
these second-order factors could not be identified with only one loading, thus the
CHC model here effectively consisted only of first-order factors loading upon a
second-order g factor. The names of the second-order factors are included in Figure
3.3 only for illustration purposes as they contributed nothing to the estimation of the
model.
In the course of fitting the VPR model, a negative residual was encountered
for the second-order Verbal factor, which was formed from the first-order factors
Information, English/Math and Math (in males z = -9.08 to -5.92, p < .001; in
females this residual was either negative or non-significant, z = -1.48 to 5.80). This
suggested that too much of the test variance was being assigned to the Verbal factor,
possibly because of the large number of tests forming the factors composing it (see
Figure 3.1). In order to resolve this issue, the Information factor was placed on its
own second-order factor (also named the Information factor), which eliminated the
negative residual and improved model fit. Based on the exploratory analysis it was
apparent that the first-order Speed factor, which consisted of mainly clerical-type
speed tests, might load onto the Verbal factor formed by the Math and English tests;
the first-order Speed factor’s highest correlation was with the English/Math factor
(see Table 3.5). Vernon (1961) observed that factors formed by clerical-type tests
often loaded within the verbal-educational domain. As predicted, placing the Speed
factor on the Verbal factor improved the fit of the model compared to the Speed
factor having a separate loading on g, and this model was used as the final version of
the VPR model.
The VPR, CHC and Gf-Gc models fit well in both males and females samples
according to conventional fit criteria (Hu & Bentler, 1999): the RMSEA was below
.060 in all cases, the SRMR below .080, and the CFI was close to .950 or greater
(Table 3.7). The VPR model demonstrated the best fit in all the samples according to
all fit indices that were used, followed primarily by the CHC model. A BIC
difference of 10 is normally considered very strong evidence in favour of the model
48
with the smaller value (A. E. Raftery, 1995); however, because of the large PT
sample sizes the saBIC values for our models were so large that a difference of 10
was trivial. Therefore, we calculated what the saBIC differences would have been if
the samples were a more conventional 500.7 As shown in Table 3.7, the VPR model
had the lowest saBIC at a sample size of 500 compared with the CHC and Gf-Gc
models in each sample. In males, the VPR had a saBIC 55.0 – 93.6 lower than the
CHC model, and 61.9 – 100.3 lower than the Gf-Gc model. In females, the VPR had
a saBIC 36.9 – 119.3 lower than the CHC model, and 43.4 – 115.10 lower than the
Gf-Gc model. The VPR model had the lowest saBIC despite being the least
parsimonious model (it containined the highest number of freely-estimated
parameters). Thus, the VPR model was found to consistently have the best fit to the
broad selection of PT tests. The CHC model had a lower saBIC than the Gf-Gc
model in five out of eight samples, but the saBIC difference was lower than 10 in all
cases, thus in general the difference in fit between these models was marginal.
7 The formula for the saBIC is -2(log-likelihood) + pln[(N +2)/24], where p is the number of freely-
estimated parameters. Since log-likelihood is linearly related to sample size, it was scaled by the ratio
of 500 to the full sample size for each sample. In the second half of the equation, 500 was entered for
N.
49
Table 3.7
Fit statistics of confirmatory factor models for the broad selection of PT tests.
Sample Sample
size
χ2 df saBIC RMSEA
(95% CI)
CFI SRMR saBIC,
sample
of 500
Males
VPR model
Grade 9 49264 45592.64 592 8755769.82 .039 (.039-.040) .955 .036 89301.38
Grade 10 48561 51421.27 593 8761767.73 .042 (.042-.042) .951 .037 90652.44
Grade 11 44172 51480.47 594 8015636.24 .044 (.044-.044) .949 .036 91163.59
Grade 12 38894 47517.10 592 7040873.99 .045 (.045-.045) .948 .036 90949.57
Cattell-Horn-Carroll Model
Grade 9 49264 52812.82 599 8762936.61 .042 (.042-.042) .948 .039 89356.34
Grade 10 48561 59600.54 599 8769901.32 .045 (.045-.045) .943 .040 90715.45
Grade 11 44172 61253.79 600 8025364.45 .048 (.048-.048) .939 .040 91255.98
Grade 12 38894 56212.83 598 7049525.37 .049 (.049-.049) .938 .041 91043.11
Fluid-Crystallized model
Grade 9 49264 52597.02 596 8762743.69 .042 (.042-.042) .948 .038 89363.28
Grade 10 48561 59357.72 596 8769681.34 .045 (.045-.045) .943 .040 90722.07
Grade 11 44172 60878.92 597 8025012.13 .048 (.047-.048) .939 .040 91263.89
Grade 12 38894 55655.29 595 7048990.01 .049 (.048-.049) .939 .040 91045.06
Females
VPR model
Grade 9 49973 36559.98 595 8863188.52 .035 (.034-.035) .962 .025 90385.69
Grade 10 48237 37640.05 591 8598498.30 .036 (.036-.036) .961 .025 88974.32
Grade 11 46504 40765.08 591 8329464.47 .038 (.038-.039) .958 .032 94724.80
Grade 12 41119 42975.16 594 7378914.32 .042 (.041-.042) .951 .042 95289.34
Cattell-Horn-Carroll Model
Grade 9 49973 41993.83 601 8868576.52 .037 (.037-.037) .956 .026 90422.60
Grade 10 48237 44907.56 597 8605720.18 .039 (.039-.040) .954 .027 89030.91
Grade 11 46504 50429.74 597 8339083.71 .042 (.042-.043) .948 .034 94815.95
Grade 12 41119 53673.68 600 7389568.16 .046 (.046-.047) .938 .037 95408.62
Fluid-Crystallized model
Grade 9 49973 41739.25 598 8868344.86 .037 (.037-.037) .956 .026 90429.14
Grade 10 48237 44481.76 594 8605317.19 .039 (.039-.039) .954 .026 89035.65
Grade 11 46504 49579.24 594 8338255.94 .042 (.042-.042) .949 .033 94815.45
Grade 12 41119 52638.67 597 7388555.49 .046 (.046-.046) .939 .036 95404.44
3.3.2.2 Narrow selection
Confirmatory factor models for the narrow selection of tests were developed
in the same manner as those for the broad selection, based upon the exploratory
analysis (see Table 3.8 for the first-order loadings; differences in the loadings across
each sample were detailed in 3.1.2.). Once again cross-loadings for Math tests onto
the English achievement (in the CHC model) and Verbal comprehension factor (in
the Gf-Gc model) were fixed to zero, as was the loading of Word Functions in
Sentences on their Math factors.
50
The initial VPR model fit properly in all grades except in grade 9 males,
where a small negative residual was encountered for the Math factor (standardized
loading = -.034). Fixing this residual to zero resulted in another negative residual for
the Math 2 test, and a negative loading for this test on the English/Math factor
(standardized loading = - .35), indicating that this factor loading was the source of
model misfit. Thus, the Math 2 loading was removed from English/Math, which
resolved the negative residual.
The second-order structures of the models were the same as in the broad
selection, except the absence of a first-order Mechanical/Science factor meant that
there was no longer a second loading onto the Perceptual factor in the VPR model or
a third loading on the Gc factor in the Gf-Gc model. In the CHC model, the second-
order factor for science knowledge (K1) was absent. The Verbal factor occupied the
same role as the Information factor did in the broad selection.
Table 3.9 contains the fit statistics for the models based on the narrow
selection. The VPR model again fit the data best according to all criteria (except for
two comparisons where the SRMRs were equal). The second best-fitting model was
generally the CHC model, but the saBIC differences between the CHC model and
Gf-Gc model were again small and inconsistent, pointing to only a marginal
difference overall between them. In the hypothetical male samples of 500, the VPR
had a saBIC 8.4 – 17.5 lower than the CHC model, and 14.8 – 24.6 lower than the
Gf-Gc model. In females, the VPR had a saBIC 15.8 – 34.4 lower than the CHC
model, and 21.3 – 30.6 lower than the Gf-Gc model. Although the saBIC difference
between the VPR and CHC model was less than 10 in grade 9 males, a difference of
8.4 is still characterized as “strong” evidence of a model fit difference, and
corresponds to a posterior odds of 66:1 in favor of the VPR model (Raftery, 1995).
The AIC, RMSEA and CFI also favored the VPR over the CHC model. Thus, the
VPR model demonstrated the overall best fit to the narrow selection of PT tests.
51
Table 3.8 First-order loadings for the CHC, Extended Gf-Gc and VPR models in the narrow selection (grade 10 males). Test Primary loadings Secondary loading(s)
CHCa
Gf-Gcb
VPR CHC Gf-Gc VPR
Vocabulary K0 Vi Verbal A6 V English/Math Memory for sentences A6 V English/Math Memory for words A6 V English/Math Disguised words A6 V English/Math K0, P Vi,P Verbal, Speed Spelling A6 V English/Math Capitalization A6 V English/Math Punctuation A6 V English/Math English usage A6 V English/Math Effective expression A6 V English/Math Word functions in sent. A6 V English/Math Math
Reading comprehension A6 V English/Math K0 Vi Verbal Creativity K0 Vi Verbal Vz,
A6 Visualization, V
Spatial/Reasoning, English/Math
Mechanical reasoning Vz Visualization Spatial/Reasoning K0 Vi Verbal Visualization in 2D Vz Visualization Spatial/Reasoning Visualization in 3D Vz Visualization Spatial/Reasoning Abstract reasoning Vz Visualization Spatial/Reasoning A6 V English/Math Math 1 KM Gq Math P P English/Math, Speed Math 2 KM Gq Math English/Math
Arithmetic comp. KM Gq Math P P Speed, English/Math Table reading P P Speed Clerical checking P P Speed Object inspection P P Speed Vz Visualization Spatial/Reasoning
a K0 = general verbal information, A6 = English achievement, KM = math achievement, P = perceptual speed.
b Vi = general information, V = verbal comprehension, Gq = quantitative knowledge.
52
Table 3.9
Fit statistics of confirmatory factor models for the narrow selection of PT tests.
Sample Sample
size
χ2 df saBIC RMSEA
(95% CI)
CFI SR
MR
saBIC in
sample of
500
Males
VPR model
Grade 9 49096 18297.50 186 5724268.66 .045(.044-.045) .962 .041 58560.39
Grade 10 48254 18867.29 185 5743450.08 .046(.045-.046) .962 .040 59448.13
Grade 11 44165 17562.76 184 5254414.22 .046(.046-.047) .963 .037 59755.12
Grade 12 38877 16526.62 184 4611559.86 .048(.047-.048) .962 .037 59577.66
Cattell-Horn-Carroll Model
Grade 9 49096 20313.12 190 5726253.80 .046(.046-.047) .958 .041 58568.75
Grade 10 48254 21446.22 190 5745990.95 .048(.047-.049) .957 .041 59459.50
Grade 11 44165 20453.84 189 5257267.72 .049(.049-.050) .957 .039 59772.65
Grade 12 38877 18794.95 189 4613791.25 .050(.050-.051) .956 .039 59591.63
Fluid-Crystallized model
Grade 9 49096 20045.70 187 5726009.25 .047(.046-.047) .959 .041 58575.15
Grade 10 48254 21272.96 187 5745840.53 .048(.048-.049) .957 .041 59466.84
Grade 11 44165 20276.00 186 5257112.43 .049(.049-.050) .957 .038 59779.76
Grade 12 38877 18555.20 186 4613573.67 .050(.050-.051) .957 .038 59597.67
Females
VPR model
Grade 9 49922 16921.93 185 5815009.26 .043(.042-.042) .967 .037 59487.47
Grade 10 48209 16689.62 184 5648320.43 .043(.043-.044) .968 .036 58470.86
Grade 11 46499 17314.84 187 5468164.36 .044(.044-.045) .965 .035 62166.16
Grade 12 41095 16920.67 187 4830928.59 .046(.046-.047) .961 .034 62390.14
Cattell-Horn-Carroll Model
Grade 9 49922 20269.31 191 5818310.80 .045(.045-.046) .960 .038 59503.31
Grade 10 48209 19909.80 189 5651502.59 .047(.046-.047) .961 .038 58488.84
Grade 11 46499 21025.72 191 5471844.96 .048(.048-.049) .958 .037 62196.01
Grade 12 41095 21537.53 191 4834515.66 .051(.050-.052) .953 .037 62424.49
Fluid-Crystallized model
Grade 9 49922 19912.12 188 5817976.53 .046(.045-.046) .961 .038 59508.80
Grade 10 48209 19487.37 186 5651102.98 .046(.046-.047) .962 .037 58493.61
Grade 11 46499 20282.65 188 5471124.61 .048(.047-.048) .959 .035 62196.72
Grade 12 41095 18647.02 188 4832647.49 .049(.048-.049) .957 .033 62409.30
3.3 Discussion
We found that the VPR model provided the best fit to the Project TALENT
battery of the three models that were tested. Three studies now support the
conclusion that the VPR model provides a better description of the structure of
intelligence than the CHC model or its precursor the three-stratum model (Johnson &
Bouchard, 2005b, Johnson, te Nijenhuis, & Bouchard, 2007), and four studies now
support the VPR model over the Gf-Gc model (Johnson & Bouchard, 2005a, 2005b;
53
Johnson, Te Nijenhuis, et al., 2007). No study to date has contradicted this
conclusion.
Nonetheless, the PT dataset had some limitations for testing the structure of
intelligence differences. As always, test selection was of relevance to how the
different theoretical models were represented. The PT test battery was not suited to
testing all of the differences between the models because it lacked tests in certain
domains such as long-term storage and retrieval (Glr) and reaction and decision
speed (Gt) (McGrew, 2009). However, in its favor, the PT battery was not
constructed according to any particular theory of intelligence, and, as we argue
below, it was suitable for testing each model’s predictions in mechanical/scientific
and verbal/educational domains. A second potential limitation of our findings is that
even though the sample was representative of American high school students in the
early 1960s, the pattern of results may not generalize to more recent samples due to
cultural and educational changes. The structure of ability may also shift with age,
becoming more differentiated as students have the opportunity to learn more
specialized knowledge and subsequently enter the workforce as adults (Li et al.,
2004).In this section, we first discuss some possible reasons why the VPR theory
outperformed CHC/Gf-Gc theory in explaining the structure of the PT battery.
Second, we describe some differences in how the VPR model was manifested in PT
compared to previous studies. Third, we turn to the theoretical implications of this
study and previous comparison studies of the structure of intelligence.
3.3.1 The three theories in Project TALENT
The VPR model described the overall structure of the PT battery more
accurately than the CHC and Gf-Gc models. Yet one possible disadvantage of the
battery for the CHC model may have been that it effectively lacked second-order
factors (such as Gc, Grw, etc.; see Figure 3.3) because there were not enough
indicators for them. Nevertheless, CHC theory dictates that the six first-order factors
we observed all load onto different second-order factors; thus the creation of second-
order factor that were combinations of different first-order factors would not have
been consistent with the theory (McGrew, 2009). In addition, such a model with
only one higher-order factor received support from Carroll (1993), who found this
54
structure in the male-only PT datasets SHAY01 and FLAN01. Such a model has
been used as representative for CHC theory in PT by later researchers (Reeve, 2004;
Reeve et al., 2006) Although Carroll (1993) found three second-order factors in the
female-only dataset SHAY02 (see Tables 15.4 and 15.14, Carroll, 1993), two of
these factors were inconsistent with both his three-stratum theory and the more
recent CHC theory. These factors were the 2H factor (the combination of fluid and
crystallized intelligence), which is lacking in both models, as is the Technical
Information factor that received loadings from the General Information factor (K0),
Math knowledge (KM) and General Science Information (K1). Thus, our version of
the CHC model was consistent with the structure Carroll found in the male PT
datasets, and was true to the most recent description of the CHC model in terms of
the classification of first-order factors as loading onto particular second-order factors
(McGrew, 2009). Nevertheless, our results contradicted the prediction derived from
CHC theory that the six first-order factors in PT belong to six different domains (and
thus make six independent contributions to g; see Figure 3.3). Similarly, our results
also contradicted the four second-order factors of the Extended Gf-Gc model:
crystallized intelligence (Gc), visualization (Gv), quantitative abilities (Gq) and
cognitive speed (Gs); see Figure 3.2. Some of the second-order factors of the Gf-Gc
and CHC models were likely not fully representative of the broad abilities in the
models due to the relative narrowness of the PT test battery in certain domains (such
Gs or Gq), but if the structure of ability were divided into those domains, rather than
those predicted by the VPR model, then the models should still have demonstrated
better fit than the VPR model. These same arguments can also be extended to the
three models in the narrow selection of tests.
As noted above, we believe that, despite limitations for testing these theories
in terms of the test battery content, PT was suitable for testing their predictions about
tests in the mechanical/scientific and verbal/educational domains. An important area
where the VPR model and CHC/Gf-Gc models differ is in their conceptualizations of
where tests of mechanical/science knowledge fit in the structure of intelligence. The
VPR model predicts that performance on these tests depends on the underlying
Perceptual and/or Image Rotation abilities and skills of the test taker, as well as
interest in and experience with the subject area. Thus the VPR model specifies that
55
tests of mechanics and science information are distinct from measures of acculturated
knowledge that rely more directly on verbally expressed knowledge (e.g. General
Information). The Gf-Gc model does not make this distinction and instead holds that
all measures of acquired knowledge are part of crystallized intelligence (Gc) as
distinct from non-learned reasoning and information processing capacities. Although
the CHC model does distinguish factors of domain-specific knowledge from Gc, it
does not place specific mechanical/science knowledge with spatial ability as the VPR
model does. The PT battery provided a difficult test for the VPR model because the
mechanical and scientific information tests were based entirely on the recall of stored
knowledge, and the Spatial/Reasoning factor was made up of tests that were nearly
all visual-spatial and based on novel problem-solving. The results of the current
study thus provide support for the VPR view that spatial ability and
mechanical/science knowledge are linked under Perceptual ability.
Another area where the models differ is on how tests of English and Math
ability fit into the intelligence hierarchy. Both the Gf-Gc and CHC models dictate
that Quantitative abilities and English language abilities load onto separate second-
order factors (Horn & Blankson, 2005; McGrew, 2009). The VPR model, in
contrast, proposes that Math and English abilities are linked under the broad Verbal
factor (though Math abilities also often load on the Perceptual factor, see Johnson &
Bouchard, 2005b). Thus, the VPR model can also explain why a first-order
English/Math factor was obtained in the PT battery, but the Gf-Gc and CHC model
cannot.
Despite favoring the VPR model, the difference in fit according to the CFI
and SRMR were not large, particularly for the narrow selection. The interpretation
of the VPR as better fitting was dependent on the validity of a handful of cross-
loadings. It is possible that proponents of the Gf-Gc and CHC theories could
propose alternative better-fitting specifications of their models in Project TALENT.
The VPR model also benefitted from a more flexible theoretical stance (see section
3.3.3 for further discussion).
In summary, the better fit of the VPR model may be explained because it
made allowance for the underlying relation of the Spatial ability and
56
Mechanical/Science factors in the form of the second-order Perceptual factor, and the
relation between the English and Math factors in form of the second-order Verbal
factor. The lack of these factors in the CHC/Gf-Gc models and their too precise
specification of second-order factors likely explain their poorer fits to the data.
3.3.2 Variations in VPR model specifications
The specifications of the VPR model in the studies that have fit it have
varied. One notable difference in the PT VPR models from previous ones was that
model fit was improved when the Information factor was separated from the broad
second-order Verbal factor. Previous studies have only found only one Verbal factor
in the VPR model (Johnson & Bouchard, 2005a, 2005b; Johnson, Te Nijenhuis, et
al., 2007). Separation of the two factors may have been helpful because there was a
surplus of Information and English/Math-related tests in this school-oriented battery,
creating a spuriously high correlation between g and the Verbal factor with the
Information loading upon it. This suggests that these two factors were bloated
specific (overrepresentation of tests in closely related subject areas), as indicated by
the high number of tests loading upon them relative to the other factors (see Tables
3.5 and 3.8). At the same time, the finding of more than one second-order Verbal
factor is not inconsistent with VPR theory, as, like Vernon (1961), it proposes that
the number of factors at any stratum and the precise borders between them are
functions of the specificity of the tests in the battery rather than inherent facts of
nature (see section 3.3. below for further discussion).
In the first study on the VPR model, the first-order Perceptual Speed factor
was subsumed under the second-order Perceptual factor (Johnson & Bouchard,
2005b). Here, however, the first-order Speed factor was found to correlate highest
with the English/Math factor, and thus was placed under the Verbal factor. These
findings support Vernon’s (1961) view that speed factors can be subsumed by either
of the Verbal or Perceptual factors, depending on the overall battery test content. In
the PT test battery, the Speed factor was largely made up of tests with
verbal/numerical content (e.g. Table Reading, Clerical Checking, Disguised Words).
This factor can thus be compared more closely to the Fluency factor found in the
operationalization of the VPR model in the Thurstone and Thurstone (1941) test
57
battery (Johnson and Bouchard, 2005b). In the VPR model specified in Wolff and
Buiten's (1963)’s battery, in contrast, the Speed factor was more perceptual in nature,
involving more visual than verbal/numerical stimuli (Johnson, te Nijenhuis, &
Bouchard, 2007). Despite its loading on the Verbal factor, the Speed factor in the
current study did potentially also have a Perceptual component, as illustrated by the
cross-loading of the Object Inspection test on the Spatial/Reasoning factor (see
Tables 3.2 and 3.3). Thus, different factors that are each labelled as speed across
studies may have quite different characteristics depending on the content of the tests
involved in the factor.
The factors underlying the Perceptual factor in the PT battery can be
compared most closely with those obtained in the analysis of de Wolff and Buiten
(1963)’s test battery (Johnson, te Nijenhuis & Bouchard, 2007). In that battery, a
first-order factor of mechanical reasoning was formed from tests of knowledge about
tools and reasoning using mechanical principles, which is comparable to the
Mechanical/Science factor we obtained in the broad selection of PT tests. Similar
factors of mechanical reasoning were not identified in the two other studies on the
VPR model because these test batteries lacked tests in that domain, with the
exception of the mechanical ability test in the Minnesota Study of Twins Reared
Apart (MISTRA) battery, which loaded onto the spatial factor (Johnson & Bouchard,
2005b). A second similarity with the Perceptual factor in de Wolff and Buiten
(1963)’s battery is that its highest loading was from a factor of inductive reasoning
formed mainly by matrix reasoning tests, which is similar to the Spatial/Reasoning
factor of the current study. One remaining difference between the VPR model in the
PT and previous studies is the lack of a separate Image Rotation factor, but this can
be explained by the deficiency of pure tests for this ability in the PT battery aside
from the Visualization in Three Dimensions test.
3.3.3 Theoretical implications for the structure of intelligence
An important aspect of VPR theory that distinguishes it from other theories of
intelligence is the proposal that the main dimension, after g, along which the
structure of intelligence is organized is the Verbal-Perceptual-Image Rotation
dimension. Johnson and Bouchard (2005a) postulated that this dimension reflects
58
the fact that Verbal abilities rely on the serial, analytic functions of the left brain
hemisphere, while Perceptual and Image Rotation abilities rely more on the non-
verbal, parallel functioning of the right hemisphere (see also Gustafsson, 1984). In
further support of this view, Johnson and Bouchard (2007) found that when g was
partialed from from the 42-test MISTRA battery, there was a residual dimension with
Verbal abilities on one pole and Image Rotation and Perceptual abilities on the other.
This dimension also displayed a strong sex difference. Further research has
suggested that position on the Verbal-Image Rotation dimension is associated with
regional brain differences (Johnson, Jung, Colom, & Haier, 2008).
McGrew (2009) observed that while the CHC framework has already
had a strong influence on applied fields, “the adoption of the CHC umbrella term has
been much slower in theoretical fields, such as research published in the journal
Intelligence.” (p. 3). One possible reason that basic researchers have been slow to
embrace it is that, in contrast with VPR theory, CHC theory lacks the theoretical
content and predictions to satisfy researchers’ requirements for a basic theory of
intelligence. Carroll (1993) stated that a theory about the structure of cognitive
abilities should “provide hypotheses about the sources of individual differences in
these abilities” (p. 631); however, he proposed or tested few such hypotheses for his
three-stratum theory, and subsequent CHC researchers, as well as Extended Gf-Gc
researchers, have seemed satisfied to produce ever-growing catalogues of ability
factors, without substantial theoretical investigation into the cognitive, genetic or
neurological underpinnings of these new factors (McGrew, 2009; Horn & Blankson,
2005).
One strong theoretical prediction of VPR theory, derived from its dimension
view of intelligence, is that any attempt to specify the number of factors at any order
or stratum is inappropriate because the dimension is continuous. The implication of
this view, which is different from the discrete-factors view of the CHC and Gf-Gc
theories, is that there are many points along the Verbal-Image Rotation dimension
where factors could be found if appropriate tests were used. Thus, the only
prediction for the content of factors is that they lie somewhere along this dimension,
with their specific character dependent on the content of the tests in the battery.
59
Nonetheless, VPR theory would not predict, for example, that a factor would be
found which combines Verbal and Image Rotation tests, because these represent
positions that are far apart on the dimension. Another related prediction of VPR
theory is that even the number of strata is dependent on the level of detail of
measurement, and thus the number of strata could potentially be increased
indefinitely with tests of sufficient breadth and detail. These theoretical points of
view contrast with those in the Gf-Gc and CHC theories that the structure of
intelligence is organized into a limited number of identifiable factors and strata.
Another important theoretical difference is that in the CHC and Gf-Gc
models the broad second-order factors are conceptualized to depend more on either
purportedly novel processing abilities or acquired knowledge (Carroll, 1993;
McGrew, 2009; Horn & Blankson, 2005), whereas the VPR model maintains that
Verbal, Perceptual and Image Rotations abilities all involve stored knowledge, which
includes knowledge about how to approach problems involving freshly-generated
processing, with the balance depending on the particular constructions of the tests
and the experiences and knowledge of the test takers. For example, the Perceptual
factor in the PT VPR model is incongruent with CHC and Gf-Gc theory because it
received loadings from a factor made up from information tests
(Mechanical/Science), and a factor made up of (often-presumed novel) reasoning
tests (Spatial/Reasoning), but VPR theory holds that these factors are linked because
of their reliance on similar spatial/mechanical content and underlying perceptual
ability. This concept seems to have been overlooked in the literature, and in test
design, because the distinction between crystallized and fluid factors is often
confounded with the division between tests involving verbal and spatial content.
Carroll (1993)’s classifications are also subject to this criticism. Of the factors
classified as Gf in his review, the three most frequent first-order factors that had high
loadings on Gf were Induction, Visualization and Sequential Reasoning (Carroll,
1993, p. 598). Induction was defined most often by Ravens Progressive Matrices,
Visualization was made up entirely of loadings from visual-spatial tests, and
Sequential Reasoning was defined most often by a test called Ship Destination,
which involved simple calculations based on a diagram of ship locations (Table 6.1,
Carroll, 1993). The first-order factors that Carroll (1993) most frequently proposed
60
to load onto Gc were Verbal Ability, Language Development, and Reading
Comprehension, all factors which were dominated by verbal subject-matter content
(p. 599). A more recent example of the conflation of spatial and verbal factors with
fluid and crystallized factors is given by Benson, Hulac and Kranzler (2010).
Benson et al. (2010) found that the CHC model provided a better fit to the Fourth
Edition of the Weschler Adult Intelligence Scale (WAIS-IV) than the structure
proposed by the test designers. However, in the CHC model the Gc factor was
defined solely by loadings from tests of verbal subject-matter, and the Gf factor
received loadings from only visual-spatial tests plus the Arithmetic test.
In this study, we found again that the VPR model, which contains factors of
verbal content and perceptual-spatial content that transcend the fluid/crystallized
distinction, yielded better fit than models specifying a separate crystallized factors
(such as Gc and Gkn) and fluid factors (such as Gv and Gs). This finding thus
provides further evidence that the higher-order structure of intelligence is organized
along a Verbal-Perceptual-Image Rotation dimension rather than characterized by
broad factors distinguished mainly by their purported reliance on novel processing or
acquired knowledge.
3.4 Linking cognitive ability with personality
Establishing the VPR model as the best-fitting model among the three most
supported in the literature enabled exploration of the associations among cognitive
ability and personality and occupational interests, using the VPR model as the basis
for consideration of cognitive ability. The first step in this process was to examine
personality-intelligence associations.
The PT data provided an opportunity to look not just at linear associations but
also nonlinear associations, which require a large sample due to their relatively small
effect sizes. Some research in the gifted literature provided a basis for hypothesizing
that such nonlinear associations could be substantive at the right tail of the
intelligence distribution. Hypotheses for the linear associations were drawn
primarily from reviews of previous personality-intelligence research.
61
The analysis did not include specific abilities because there have been fewer
studies of specific abilities and personality, hence making hypotheses in this area
more difficult to formulate. A second reason was to keep the number of analyses
manageable. As in the first study, all eight PT grade and gender samples were used.
62
Chapter 4: Linear and nonlinear associations between general intelligence and personality
4.1 Introduction
Intelligence and personality are important predictors of behavior and
outcomes in many domains, notably in educational and occupational settings
(Barrick & Mount, 1991; Hunt, 2011). In addition, there are some associations
between intelligence and personality traits (Ackerman & Heggestad, 1997;
DeYoung, 2011). Within the Big Five framework, general intelligence (g) is most
strongly associated with Openness to Experience (r = .33 in the N-weighted meta-
analysis of Ackerman & Heggestad, 1997). This connection may seem obvious since
measures of Openness to Experience typically include items assessing engagement in
intellectual pursuits, and because intelligence has often been held to be the cognitive
part of personality (Cattell, 1950; DeYoung, 2011; Guilford, 1959). Nonetheless,
intelligence is also related to personality traits that are considered the least cognitive,
such as Extraversion and Neuroticism (DeYoung, 2011). Neuroticism has
consistently shown modest negative correlations with general intelligence (r = -.15 in
Ackerman & Heggestad, 1997), and most recent studies (performed after the year
2000) have shown that Extraversion also has a small but significant negative
association with g, in the range of r = -.04 to -.11 (Luciano, Leisser, Wright, &
Martin, 2004; Moutafi, Furnham, & Paltiel, 2005; Soubelet & Salthouse, 2011; Wolf
& Ackerman, 2005). In addition, DeYoung (2011) found that in 9 studies not
included in Ackerman and Heggestad’s (1997) meta-analysis, Conscientiousness had
a mean N-weighted correlation of -.12 with intelligence.
Nevertheless, some researchers have argued that the theoretical implications
of these personality-intelligence correlations are limited due to their small size or
inconsistency across studies (Eysenck, 1994; Soubelet & Salthouse, 2011; Zeidner,
1995). One possibility is that some intelligence-personality associations could be
nonlinear, and thus missed by traditional linear analyses (E. J. Austin, Deary, &
Gibson, 1997; E. J. Austin et al., 2002; Eysenck & White, 1964; Reeve et al., 2006).
Findings in this area have, however, have often been negative. Austin et al. (1997)
63
found evidence for positive quadratic effects of Neuroticism and Openness to
Experience on intelligence in one sample, but Austin et al. (2002) did not find any
significant effects of this kind for the Big Five and Eysenck’s Big Three in four other
datasets. There are three theoretical and methodological issues surrounding these
results.
First, different theories make alternative causal predictions about personality-
intelligence relations. For example, Ackerman’s PPIK theory (intelligence-as-
process, personality, interests, and intelligence-as-knowledge) predicts that
intelligence becomes related to personality through cognitive investment in four trait
complexes which involve different personality traits and interests (Ackerman, 1996;
Ackerman & Beier, 2003b). Alternatively, Chamorro-Premuzic and Furnham (2006)
proposed that personality-intelligence relations can be conceptualized as the
influence of personality traits on intellectual competence, where intellectual
competence is defined as “an individual’s capacity to acquire and consolidate
knowledge throughout the life span” (p. 259, Chamorro-Premuzic & Furnham,
2006). PPIK theory suggests that cognitive factors causally contribute to broader
constellations involving personality and interests (trait complexes), and thus that the
association between all the variables is an emergent property due to reciprocal
causation between all three variables. In contrast, Chamorro-Premuzic and
Furnham’s theory proposes that personality traits directly influence the development
of intelligence. A third possibility is that intelligence contributes directly to the
development of personality through conscious perceptions of adaptive benefit of
particular behaviours, or through the influence of intelligence on motivations.
When estimating only linear effects, it is difficult to distinguish these
possibilities without a longitudinal design, because effects are typically symmetrical
no matter which ways the causal arrows are drawn. However, nonlinear analyses can
pick up larger effects in one direction (e.g. there might be a quadratic effect of
intelligence on Extraversion but no quadratic effect of Extraversion on intelligence),
which can suggest that causal forces operated in this direction. Previous studies of
quadratic effects have focused on the quadratic effects of personality on intelligence
64
(E. J. Austin et al., 1997; E. J. Austin et al., 2002); however, in the current study we
assessed quadratic effects of intelligence on personality.
The second issue surrounding nonlinear personality-intelligence relations is
that previous nonlinear studies were not performed with latent variables as predictors
but with observed scores (E. J. Austin et al., 1997; E. J. Austin et al., 2002; Reeve et
al., 2006). This limited their power because the size of the quadratic effect was not
corrected for unreliability. Quadratic terms are particularly sensitive to unreliability
of the predictor variable (Moosbrugger, Schermelleh-Engel, Kelava, & Klein, 2009).
Methodological researchers have observed that even using factor scores for the
predictor can produce biased estimates of structural model parameters due to residual
measurement error (D. Bartholomew, 1987; Harring, Weiss, & Hsu, 2012).
Recently, Harring et al. (2012) found that, compared with methods that model latent
quadratic terms directly, the use of factor scores led to substantial underestimation of
quadratic coefficients.
A third issue is that to detect quadratic effects with small effect sizes, large
sample sizes are needed. Under simulation, Harring et al. (2012) showed that for a
medium-sized quadratic effect that accounted for 5% of the variance, even a small
sample size of 50 was sufficient to obtain power over .80. However, in practice,
quadratic or interaction effects can be considerably smaller than this, accounting for
only 1% or 2% of the variance. To find these effects, very large sample sizes (i.e. of
500 or greater) are necessary. For example, Moosbrugger et al. (2009) found that for
a quadratic effect size of 2% and a sample size of 400, average power was only 76%
using latent estimation (power would be less with non-latent methods). Thus, the
sample of Austin, Deary & Gibson (1997) and two of the four samples in Austin et
al. (2002) may not have had sufficient power to detect small quadratic effects.
Reeve, Meyer and Bonaccio (2006) conducted one study on personality-
intelligence relations that was sufficiently powered. Their study is directly relevant to
ours as we made use of the same sample so we review their analysis in detail. Reeve
et al. (2006) used a subsample of data from Project TALENT (PT), a nationally-
representative study of approximately 400,000 American high school students in
1960. The sample in Reeve et al. (2006) consisted of 71,887 students in their final
65
year of high school (seniors), with a mean of age 17.2 years (SD = 1.3). The ten PT
personality scales were developed specifically for the Project in the late 1950s,
before there was much consensus about models of personality structure. The scales
used thus do not correspond directly to the Big Five framework in common usage
today, but Reeve et al. (2006) related the scales to the Big Five by two methods.
First, the three authors independently examined each scale’s content and compared it
to the content of the NEO-PI-R scales (P.T. Costa & MacCrae, 1992), and second,
they re-administered the PT personality scales and IPIP scales for the Big Five to a
sample of 219 college students. Table 4.1 summarizes the NEO-PI-R facet with
which each PT scale was most closely associated (by rater consensus), as well as
with which Big Five trait(s) the scales loaded in a joint factor analysis with IPIP
scales (Reeve et al., 2006).
These relations provided a way to link the PT scales to the larger literature on
personality-intelligence relations, which has frequently been organized according to
the Big Five (e.g., Ackerman & Heggestad, 1997; Austin et al., 2002). The facet-
matching by Reeve and colleagues may be limited due to imperfect content overlap,
but the majority of the PT scales displayed good convergent validity with the Big
Five factors predicted to subsume them (factor loadings = .42 to .81). In addition,
the content of the PT scales was facet-like; hence they could be viewed as analogous
to facets of the Big Five, with the exception that some scales (e.g. Self-Confidence)
would be facets of more than one Big Five factor.
Table 4.1 Associations of the Project TALENT personality scales with the Big Five. PT scale NEO-PI-R facet Big Five trait loading(s)
a Sociability Gregariousness (E) E (0.69), A (0.38) Calmness Anger (ES) - reversed ES (0.69) Vigor Activity (E) E (0.43) Social sensitivity Sympathy (A) A (0.81) Tidiness Orderliness (C) C (0.79) Culture Aesthetics (O) O (0.51) A (0.44) Self-confidence Self-consciousness (ES) -
reversed E (0.60) ES (0.60)
Mature personality Achievement Striving (C) C (0.63) A (0.35) Impulsiveness Cautiousness (C) - reversed E (0.42) Leadership Assertiveness (E) E (0.51) O (0.41) a Loadings obtained in a joint factor analysis of the IPIP and PT scales by Reeve et al. (2006)
66
Reeve et al. (2006) found that g correlated positively and substantively
(above .15 in their definition) with the scales Mature Personality, Calmness, and
Self-Confidence in grade 12 males. These correlations were also observed in grade
12 females, where Culture and Social Sensitivity were also correlated positively with
g. These associations may, however, have been influenced by measurement artefacts
because the PT personality scales were nearly uniformly positively correlated with
each other. The mean of the inter-scale correlations in the senior sample was .38 in
males, and .35 in females (SD = .14 in both samples). Reeve et al. (2006) did not
address this common variance among personality scales (similar factors in other
personality inventories have been termed ‘general factors of personality’; Rushton &
Irwing (2008). This common variance was relevant because it correlated positively
with g in Project TALENT (mean r = .28 in all samples), and thus we predicted that
it would affect the correlations of the scales with g.
Recent research has suggested that the common variance between Big Five
measures is in large part due to rater bias. In a meta-analysis of 45 multi-trait multi-
method samples, Chang, Connelly and Geeza (2012) found that much of the common
variance between Big Five personality scales is due to method variance specific to
raters, which likely includes response biases such as socially desirable responding.
After rater effects were controlled for in the CTOM (correlated traits, orthogonal
methods) model, adding a general factor of personality (GFP) above the Big Five
factors resulted in a substantial decrement in model fit compared the model allowing
free covariance between the Big Five (Chang et al. 2012). Moreover, the GFP had
non-substantive loadings from Extraversion (.03) and Openness to Experience (-.09),
supporting the view that there is no single factor that sits above the Big Five in multi-
informant data (however, a model with Digman’s Alpha and Beta were still found to
be plausible by Chang et al., 2012). A number of studies have now supported the
conclusion that the GFP emerges for artifactual methodological reasons (Anusic,
Schimmack, Pinkus, & Lockwood, 2009; M.C. Ashton, Lee, Goldberg, & de Vries,
2009; Bäckström, Björklund, & Larsson, 2009; de Vries, 2011).
As detailed further below, we also observed that several of the PT scales in
Reeve et al. (2006), and in our initial analysis, displayed stronger positive
67
correlations with g than were expected based on such correlations in the Big Five,
with which the PT scales correlated. This, combined with the moderate correlation
of the GFP to g, suggested that common variance may have acted as a confounder in
the estimates. Because we were primarily interested in the relations of the individual
scales to g, and wished to err on the side of under-estimation rather than over-
estimation, we conducted our analyses while controlling for the GFP.
In addition to linear associations, Reeve et al. (2006) looked for nonlinear
relations by converting the personality scores into extremeness scores (X−Meanx|)
and examining their correlations with g factor scores. Reeve et al. (2006) did not
observe any correlations between the extremeness scores and g above a selected cut-
off of .15. However, there were two limitations to their method of looking for
quadratic effects. First, whereas extremeness scores may suggest the presence of
quadratic trends, they are not equivalent to examining true quadratic effects which
predict scores with the form |X2-Meanx|. Second, Reeve et al. (2006) chose to
convert the personality scale scores rather than the intelligence test scores in PT to
extremeness scores, thus examining the effect of extreme personality on intelligence
(Rosenthal & Rosnow, 1991). This is the same direction of effect investigated by
Austin and colleagues (Austin et al., 1997; Austin et al., 2002). As noted, we were
instead interested in examining the effects of intelligence on personality. This had the
added advantage of greater power, due to the greater reliability of the latent g factor
compared to the observed personality scales.
The aim of our study was to re-examine linear and nonlinear relations between g
and personality in Project TALENT, taking into account common variance among
the scales. Moreover, we used structural equation modeling (SEM), which avoids
using factor scores and allows for direct estimation of latent linear and quadratic
effects (Klein & Moosbrugger, 2000). In addition to SEM, we used generalized
additive models (GAMs; Hastie & Tibshirani, 1986) to explore further possible
nonlinear trends. The PT data were suited to our aims because of its large and
relatively population-representative sample of nearly 400,000 high school students in
four grades, allowing for the possibility of replication across grade subsamples.
68
4.1.2 Linear personality-intelligence associations in Project TALENT
We focused on the personality-intelligence literature (primarily on the Big Five)
in generating our hypotheses about specific associations.
Openness to Experience displays a positive correlation with g (Ackerman, 2009;
Ackerman & Heggestad, 1997; DeYoung, 2011), and the two PT scales that loaded
significantly on Openness to Experience were Culture and Leadership (Reeve et al.,
2006). Neither scale is a pure measure of Openness/Intellect (see Table 4.1);
therefore, we hypothesized that their correlations with g would be positive, but
smaller in size than the .33 value in meta-analysis of Ackerman & Heggestad, 1997.
Five of the ten scales in PT had primary loadings on Extraversion, which has
typically shown small negative associations with intelligence (Wolf & Ackerman,
2005; Moutafi, Furnham & Paltiel, 2005; Austin et al., 2002). However, this relation
is not uniform for all facets of Extraversion. Ackerman and Wolf (2005) suggested
that Extraversion should be split to reflect two different aspects: social closeness (the
need for intimacy) and social potency (the need for making an impact on others).
They also hypothesized that “Individuals high on social closeness may be less likely
to invest their time in intellectually engaging tasks, leading to lower scores on
intelligence tests” (p. 533, Wolf & Ackerman, 2005). Partially consistent with this,
their meta-analysis of 48 samples showed that the correlation between social potency
and intelligence was slightly positive (r = .04, p < .05), whereas the intelligence
association with social closeness was not significantly different from zero (r = -.01)
(Wolf & Ackerman, 2005). Similarly, Pincombe, Luciano, Martin & Wright (2007)
found that the excitement-seeking and gregariousness facets of NEO Extraversion
correlated negatively with IQ (r = -.09 and r = -.15, respectively). We thus
anticipated that PT Sociability and Impulsiveness scales would show negative
associations with intelligence (due to their face-value relations with social closeness
and excitement-seeking), whereas Vigor, Self-Confidence and Leadership would
show positive associations (due to their face-value relations with social potency).
The Big Five trait Neuroticism has a negative correlation with intelligence
(Ackerman & Heggestad, 1997; DeYoung, 2011). Based on their face-value
contents, and the findings of Reeve et al. (2006), the PT scales of Calmness and Self-
69
Confidence represent the converse of Neuroticism (Emotional Stability); therefore,
we predicted these scales would display positive associations with intelligence.
The literature has suggested that Big Five Conscientiousness has a small
negative association with intelligence (Moutafi, Furnham & Paltiel, 2005; DeYoung,
2011). In addition, in a sample of British adults, Moutafi, Furnham and Crump
(2003) found that the Orderliness facet of Conscientiousness in particular had a
negative correlation with g (r = -.18), which they argued may be because lower-
intelligence individuals use planning and organization to compensate for their
disadvantage on intellectual tasks (see also Chamorro-Premuzic and Furnham, 2006).
The PT scale Tidiness was related on a content basis to Orderliness by Reeve et al.
(2006); therefore, we hypothesized that it would have a negative correlation with
intelligence. The PT scale Mature Personality was also related to Conscientiousness
by Reeve et al. (2006), hence we predicted a negative association for it.
Big Five Agreeableness has typically not been found to have significant
correlations with intelligence (Ackerman & Heggestad, 1997; DeYoung, 2011);
hence we did not make any directional hypothesis regarding the PT scale Social
Sensitivity, which was the only PT scale with a high correlation with Agreeableness
according to Reeve et al. (2006).
4.1.3 Possible nonlinear associations
Although nonlinear associations between intelligence and personality have
rarely been found, some suggestive evidence for nonlinear associations has been
found in research on gifted children and adolescents. This has primarily been done
with the Myers-Briggs Type Indicator (MBTI; Myers, McCaulley, & Most, 1985).
A meta-analysis of 14 studies of gifted adolescents, mostly identified through
talent searches using the SAT and selection into gifted programs, showed that they
were substantially more likely to fall on one side of the dichotomous MBTI
dimensions than a norm group of students (Sak, 2004). Gifted adolescents were
more likely to select Introversion over Extroversion (48.7% compared to 35.2% in
the non-gifted sample), Intuition over Sensation (71.6% compared to 31.9%), and
Perceiving over Judging (60.1% compared to 45.4%), as well as marginally more
70
likely to prefer Thinking to Feeling (53.8% compared to 47.5%; Sak, 2004). Studies
in adults have found that MBTI Extroversion to be strongly related to Big Five
Extraversion (r = .74), whereas Intuition is strongly related to Openness to
Experience (r = .72); MBTI Thinking and Perceiving are moderately negatively
correlated with Agreeableness (r = -.44) and Conscientiousness (r = - .49),
respectively (correlations for the male sample in R. McCrae & Costa, 1989).
Therefore, by extension it can be predicted that gifted adolescents may be
substantially higher in Openness to Experience and lower on Extraversion,
Agreeableness and Conscientiousness than non-gifted adolescents. A recent study of
Israeli adolescents, who were selected as the top 1% to 3% of performers on an
intelligence test, confirmed this pattern for Openness to Experience (d = .51) and
Agreeableness (d = -.28), and also showed that gifted adolescents were lower in Big
Five Neuroticism than non-gifted adolescents (d = -.26) (Zeidner & Shani-Zinovich,
2011). Group differences in Conscientiousness and Extraversion were in the
expected direction based on MBTI studies, but non-significant (Zeidner & Shani-
Zinovich, 2011).
The presence of some substantial mean differences between gifted and
non-gifted groups suggests that average personality level might differ to an
expanding (e.g. exponential) degree with increasing ability level, although it is
possible that linear effects could produce these effects as well. Exponential functions
may most closely approximate differences in certain personality traits with
increasing ability level, but such trends would also be captured by quadratic effects,
at least for one side of the parabolic curve. One issue relating to this testing is that
some studies have also found increases in personality variance with higher
intelligence (e.g. in the MBTI; Myers & McCaulley, 1985). This may violate the
assumption of homogeneity of variance underlying generalized linear models,
although these models are robust to some level of heteroscedasticity (Tabachnick &
Fidell, 2007). It is possible that higher cognitive ability is causally linked to
increases in personality variance, as intelligence potentially facilitates more flexible
adjustment of personality to the environment; however, in this study we focused on
mean-level differences in personality.
71
Given the evidence in the gifted literature, we hypothesized that positive
quadratic trends would be observed for the PT scales associated with Openness to
Experience (Culture and Leadership) as well as Emotional Stability (Calmness and
Self-Confidence). We also predicted that negative quadratic effects (an inverted-U
shape) would be observed for the scales associated with the social closeness aspect of
Extraversion (Sociability), Agreeableness (Social Sensitivity), and Conscientiousness
(Tidiness and Mature Personality). Because less is known about personality in low-
ability groups, these predictions were based on the trend for above-average
intelligence.
4.2 Method
4.2.1 Sample
Project TALENT participants were obtained by a stratified random sample of all
public and private high schools in the United States in 1960 (Flanagan et al., 1962).
The PT dataset was thus a nationally-representative sample of approximately 5% of
the student population. The full sample consisted of 376,213 students, with
approximately 100,000 students in each grade from 9 through 12. Of the full sample,
50.13% was female. The age range was from a mean of 14.4 in grade 9 (SD = .78) to
17.3 in grade 12 (SD = .67). The full individual age range was 8 to 21.
4.2.2 Intelligence measures
The intelligence measures for the current study were selected from the PT
aptitude and achievement tests, using the broad selection of 37 tests as defined in a
previous study (for descriptions of the tests and reliabilities see Major, Johnson &
Deary, 2012; see also Flanagan et al., 1962).
The data screening methods for the intelligence tests were the same as used by
Major, Johnson and Deary (2012). Scores on the PT response credibility index,
which was based on a screening test assessing illiteracy, mental disability or an
apathetic testing attitude, were used to exclude participants who did not reach the
cut-offs set by the PT study designers, except where only mental slowness was
indicated (a low score for the number of responses on the Clerical Checking test).
Transformations were applied to three tests that displayed non-normal distributions
72
(Capitalization, English Usage and Table Reading), and cases showing severe
problems with multivariate outliers were removed (Major, Johnson & Deary, 2012).
Following data screening of the intelligence tests, total sample size was reduced to
366,939 (2.47% of the sample removed, the vast majority for low screening scores).
4.2.3 Personality measures
The PT personality scale scores were derived from 108 items that asked
students how typical certain personal attributes and behaviors were of them. Table
4.2 contains sample items for the scales; reliability coefficients from Reeve et al.
(2006) are presented due to their lack of availability from the original study. The
responses to personality items were on five-point Likert scale. The scores available
in the PT dataset were scale scores, which were obtained by assigning a score of 1 to
items where the student indicated that the item described them “extremely well” or
“quite well” (the two most affirmative responses), and a score of 0 to other responses
(“fairly well”, “slightly”, or “not very well”). The converse scoring method was
used for negatively-phrased items (Wise et al., 1979).
Table 4.2 Personality test descriptives. Scale Sample item Items Reliability
a
Sociability “I like to be with people most of the time” 12 .83 Calmness “I am usually self-controlled” 9 .81 Vigor “I am full of pep and energy” 7 .76 Social
sensitivity “I never hurt another’s feelings if I can avoid it” 9 .79
Tidiness “I like to do things systematically” 11 .85 Culture “I think culture is more important than wealth” 10 .69 Self-confidence “I am usually at ease” 12 .79 Mature
personality “I make good use of all my time” 24 .90
Impulsiveness “I usually act on the first plan that comes to
mind” 9 .69
Leadership “People naturally follow my lead” 5 .65 a Reliabilities from the sample of 219 college students in Reeve et al. (2006).
4.2.4 The general factor of personality
The raw PT personality scales displayed mean inter-correlations that ranged
from .35 in the grade 12 females (SD = .14), up to a maximum of .42 in the grade 9
males (SD = .13). Across the eight samples (four grades by two genders), the first
73
common factor accounted for a mean of 41.3% of variance (SD = 2.2%). Potential
sources for this common variance included artifactual sources such as method
variance (e.g. due to pencil-and-paper testing), acquiescence bias and socially-
desirable responding, and non-artifactual true score variance.
Although it was not possible to disentangle these sources directly, some
evidence suggested that this common variance was potentially confounding the
relations of personality scales with g. Several personality scales displayed
unexpectedly positive correlations with g. The Tidiness scale, which Reeve et al.
(2006) identified on a content basis with the Orderliness facet of Conscientiousness,
displayed a positive correlation with g in all samples (mean r = .16 in males, .10 in
females; see Appendix B). This observation contradicted the finding of Moutafi et
al. (2003) that Orderliness is negatively associated with g, and that
Conscientiousness in general in also negative related (DeYoung, 2011). Similar
inferences could be drawn for the Sociability and Impulsiveness scales, which were
predicted to have negative associations with g based on the literature, but instead
showed small positive correlations (Sociability: r = .09/.05 in males/females;
Impulsiveness: r = .03/.10 for males/females)8. We hypothesized that the positive
correlation between the GFP and g could account for these positive correlations
(mean r = .28 in both males and females).
In order to aid in the interpretation of the GFP, we performed a re-analysis of
the college sample data from Reeve et al. (2006)9. The general factor from the PT
scales, extracted through maximum likelihood estimation, explained 25.9% of the
variance in the college sample. The GFP was then correlated with the individual
items (item-level data were not available in PT). The Vigor scale was over-
represented in items that correlated most highly with the GFP: six of the seven items
assessing Vigor were in the top 10 most highly-correlated items, including the most
highly correlated item (“I am energetic”, r = .63). The Vigor scale also had the
highest loading on the GFP (.71) in the college sample. In addition to this trend, only
23 of the 108 personality items (21%) contained statements that referred to other
people’s views (e.g. “people consider me sociable”), but 8 of these items were in the
8 The correlation between Impulsiveness and g in grade 9 males was non-significant, however.
9 Data obtained through personal communication (September 18, 2012).
74
top 20 most correlated with the GFP (40%). This finding suggested that items that
primed reputational concerns were more closely tied to the GFP. In the college
sample, the GFP was most highly associated with items that seemed to tap the form
of socially-desirable responding that has been termed egoistic self-enhancement
(Paulhus & John, 1998). Nonetheless, this interpretation may not generalize entirely
to the PT sample, as indicated by differences in the GFP loadings in the two samples.
In the college sample a lower loading was seen particularly on Tidiness (.29,
compared to .70 in PT). This finding indicated that Tidiness was more integral to the
GFP in PT, suggesting that the GFP in PT had more to do with Conscientiousness
than in the college sample.
Regardless of whether the correlation between the personality factor and g
was artifactual or not, our primary interest was in the relations of the individual
scales with g. Therefore, we decided to remove influence of the common variance
from the scales. The scales were regressed onto the GFP, and residuals retained for
the further analyses. To verify that the residualization did not damage the
convergent validity of the PT scales we examined their correlations with the
predicted IPIP Big Five scales in the college sample. Compared to the mean
correlation of the unresidualized scales (r = .56, SD = .13), the mean correlation
decreased to r = .36 (SD = .19). This reduction was consistent with the high
correlation of the GFP in the PT scales with the GFP in the IPIP scales (r = .77).
When the variance in the PT GFP that was explained by the IPIP GFP was removed
(through regression) prior to using it to residualize the PT scales, there was no
reduction in the correlation between the residualized PT scales and the IPIP scales
(mean r = .56, SD = .12). Removing the PT GFP appeared to reduce the convergent
validity of the PT scales, but this seemed to be because the IPIP and PT scales shared
method or rater variance, captured by their GFPs, which inflated the initial
correlations. The results of the analyses with the original scales are presented in the
supplemental materials (Appendix B), but the focus of all further presentation is on
the residualized scales.
Following residualization for the GFP, the personality scales were screened
for normality and outliers. Outliers were capped at four standard deviations above
75
and below the mean (approximately the most extreme score expected in our
samples). The scales Impulsiveness and Leadership displayed positive skewness in
all samples, therefore a square-root transformation (with reflection) was applied to
them. Following these transformations, all personality scales displayed adequate
normality (all skewness and kurtosis z values below 0.5). In contrast to the raw
scales, the mean correlation between the residualized scales was slightly negative,
and ranged from -.096 (SD = .12) in grade 11 females to -.099 (SD = .09) in grade 9
males. Table 4.3 displays the correlations in grade 10 males and females.
76
Table 4.3 Correlations among personality scales after removal of the general personality factor (frade 10 males/females).
Sociability
Social
Sensitivity Vigor Calmness Tidiness Culture Self-
confidence Mature Personality
Impul-
siveness Leadership Sociability – .019 .182 -.160 -.160 -.183 .109 -.360 .059 -.035 Calmness -.037 – -.169 -.106 -.189 -.120 -.201 -.278 -.007 -.150 Vigor .114 -.187 – -.186 -.189 -.183 -.002 -.150 .104 .020 Social sensitivity -.133 -.106 -.165 – -.112 -.213 .056 -.164 -.130 -.164 Tidiness -.162 -.203 -.157 -.155 – -.052 -.191 -.004 -.179 -.200 Culture -.127 -.068 -.191 -.202 -.041 – -.182 -.211 -.047 -.133 Self-confidence .049 -.194 -.069 .082 -.116 -.143 – -.081 -.012 .030
Mature personality -.339 -.261 -.130 -.168 -.080 -.231 -.032 – -.085 -.040
Impulsiveness .065 -.037 .035 -.095 -.145 -.048 -.072 -.103 – .116
Leadership -.051 -.082 -.043 -.145 -.186 -.092 -.049 -.079 .118 –
Male correlations are below the diagonal, females above.
77
4.2.5 Methods of analysis
We searched for linear and nonlinear associations between g and the personality
scales in two ways. The first method was to estimate linear and quadratic effects
using latent moderated structural equation modeling (LMS; Klein & Moosbrugger,
2000). LMS directly models the quadratic term as the interaction of a latent variable
with itself (or the square of the variable), and corrects for the multivariate non-
normality of the term, making it a better method than regression (Harring, Weiss &
Hsu, 2012; Moosbrugger, Schermelleh-Engel, Kelava & Klein, 2009). LMS was
performed in Mplus 5.21.
Secondly, we ran generalized additive models (GAMs; Hastie & Tibshirani,
1986), using the R package ‘mgcv’ (Wood, 2006). A GAM is a generalized linear
model in which the linear predictor depends on unknown smooth functions of the
predictor variables. The smooth functions are represented by regression splines with
a particular basis function (for our analyses, the cubic basis was selected). The
degree of smoothing of the spline is determined by the generalized cross validation
score, which is a measure of how well the spline fits across datasets with each datum
left out in turn (see Wood, 2006, for more details). We used GAMs to explore other
possible nonlinear trends apart from quadratic trends between the personality scales
and g.
Using the LMS and GAM approaches, we estimated the effects of g on the
ten personality scales in each of eight samples divided by grade and sex. We
selected the direction of effect of g on personality because we preferred this direction
theoretically and because g was more reliably measured than the personality
variables. In addition, it was not possible to estimate latent personality traits because
of the lack of item-level data. Thus, g as a predictor allowed for LMS estimation of
quadratic effects.
The measurement model used for g was the VPR model, which has been
shown to fit well to these data (Major, Johnson & Deary, 2012). The variance
explained by each effect in the LMS models was obtained by subtracting the residual
variance of the personality scales from 1 (as the personality scales were
78
standardized). GAMs were estimated with g factor scores obtained from the VPR
model.
Missing data were handled with through direct maximum likelihood
estimation, which requires the assumption the data were missing at random (MAR).
This assumption was tenable in PT because it is unlikely that students purposely
avoided particular aptitude tests or the personality scales. In addition, only 2.3 to
3.2% of the ability test scores and 1.0 to 3.5% of the personality test scores were
missing in each sample.
4.3 Results
Tables 4.4 and 4.5 display the standardized linear and quadratic effects of g
on the residualized personality scales for males and females, respectively. Figure 4.1
(males) and Figure 4.2 (females) illustrate the predicted mean-level differences in
personality based upon the estimated linear and quadratic effects in the grade 10
samples.10
Social Sensitivity in males and Calmness in females were omitted from
the figures due to the lack of significant linear or quadratic effects.
In the male samples (Table 4.4, Figure 4.1), the largest linear effects of g
were on Sociability (beta = -.042 to -.130), Calmness (beta = .076 to .104), and Self-
Confidence (beta = .106 to .131). Substantial negative quadratic effects (R2
of
approximately 2% or greater) were observed for Sociability (beta = -.146 to -.159)
and Vigor (beta = -.107 to -.116). Positive quadratic effects were observed for
Mature Personality (beta = .099 to .119) and Leadership (beta = .106 to .124).
In the female samples (Table 4.5, Figure 4.2), the largest linear effects of g
were on Sociability (beta = -.077 to -.195), Tidiness (beta = -.064 to -.163), and
Mature Personality (beta = .075 to .140). Substantial negative quadratic effects were
seen on Sociability (beta = -.140 to -.155) and Tidiness (beta = -.064 to -.163), and a
positive quadratic effect was found on Mature Personality (beta = .075 - .140).
10
The figures are at the end of the Results section. A separate page for figure titles and captions is
provided due to lack of space.
79
Table 4.4
Standardized linear and quadratic effects of g on the personality scales (males).
Trait Linear effect Quadratic effect
Gr. 9 Gr. 10 Gr. 11 Gr.12 Gr. 9 Gr. 10 Gr. 11 Gr. 12
Sociability
Beta -.042 -.087 -.118 -.130 -.146 -.148 -.155 -.159
R2 .002 .007 .014 .017 .032 .034 .037 .039
Calmness
Beta .076 .094 .101 .104 – – – –
R2 .006 .009 .010 .011 – – – –
Vigor
Beta .053 .020 – -.022 -.114 -.116 -.107 -.110
R2 .003 .000 – .000 .019 .020 .017 .019
Social Sensitivity
Beta -.029 – .017 .020 – – – -.016
R2 .001 – .000 .000 – – – .000
Tidiness
Beta – -.020 -.058 -.086 -.072 -.069 -.084 -.088
R2 – .000 .003 .007 .008 .008 .011 .013
Culture
Beta -.087 -.073 -.082 -.067 .024 .054 .077 .093
R2 .008 .005 .007 .005 .000 .004 .008 .011
Self-Confidence
Beta .131 .107 .106 .118 – – – .026
R2 .017 .011 .011 .014 – – – .001
Mature Personality
Beta .066 .061 .070 .056 .119 .115 .113 .099
R2 .004 .004 .005 .003 .022 .021 .020 .016
Impulsiveness
Beta -.125 -.076 -.032 -.023 .031 .021 .028 .018
R2 .016 .006 .001 .000 .001 .000 .001 .000
Leadership
Beta -.116 -.094 -.048 – .124 .113 .107 .106
R2 .013 .009 .002 – .023 .022 .017 .016
Effects greater than .015 were significant at p < .001, with no adjustment for multiple testing. Non-significant
effects are not shown.
80
Of our nine hypotheses about the linear effects of g on the personality traits,
five were supported in all male samples (positive: Calmness, Self-Confidence;
negative: Sociability, Tidiness, Impulsiveness), and one more received support in
some grades (the positive effect on Vigor)11
. In females, four hypotheses were
supported (positive: Culture, Self-Confidence; negative: Sociability, Tidiness) and
two had mixed support (the positive effect on Vigor and negative effect on
Impulsiveness). The most unexpected finding was a negative linear effect of g on
Leadership. This effect, in combination with the positive quadratic effect of g on
11
The negative linear effect of g on Tidiness in grade 9 males was only significant at p < .05, and
hence would not survive correction for multiple testing. Due to the effects in the other three samples,
however, we counted this effect as significant across all samples.
Table 4.5
Standardized linear and quadratic effects of g on the personality scales (females). Trait Linear effect Quadratic effect
Gr. 9 Gr. 10 Gr. 11 Gr.12 Gr. 9 Gr. 10 Gr. 11 Gr. 12
Sociability
Beta -.077 -.115 -.161 -.195 -.153 -.155 -.149 -.140
R2 .006 .013 .026 .038 .036 .037 .036 .031
Calmness
Beta – – – – – – – –
R2 – – – – – – – –
Vigor
Beta .041 – – -.032 -.074 -.077 -.065 -.050
R2 .002 – – .001 .008 .009 .007 .004
Social Sensitivity
Beta – – .017 – -.052 -.047 -.079 -.075
R2 – – .000 – .004 .008 .010 .009
Tidiness
Beta -.064 -.105 -.137 -.163 -.092 -.100 -.106 -.120
R2 .003 .011 .019 .027 .015 .016 .018 .023
Culture
Beta .032 .066 .071 .078 .030 .042 .045 .080
R2 .001 .004 .005 .006 .001 .003 .003 .005
Self-Confidence
Beta .047 .040 .034 .055 – .022 .039 .043
R2 .002 .002 .001 .003 – .000 .003 .003
Mature Personality
Beta .075 .083 .133 .140 .162 .166 .149 .122
R2 .006 .007 .018 .020 .041 .042 .035 .024
Impulsiveness
Beta -.036 .046 .055 .075 .069 .079 .046 .046
R2 .001 .002 .003 .006 .008 .010 .003 .003
Leadership
Beta -.111 -.109 -.079 -.031 .095 .093 .096 .106
R2 .012 .012 .006 .001 .014 .013 .014 .017
Effects greater than .015 were significant at p < .001, with no adjustment for multiple testing. Non-significant
effects are not shown.
81
Leadership, resulted in the highest levels of Leadership being observed for those with
low g (see Figures 4.1 and 4.2). This finding may be less trustworthy than the others,
however, because the Leadership scale only contained five items, and displayed
borderline reliability (alpha = .65) in Reeve et al. (2006). We reserve interpretation
of the meaning of the effects for the Discussion.
Of our eight predicted quadratic effects, four were supported in male samples
(positive: Culture, Leadership; negative: Sociability, Tidiness), and six were
supported in female samples (positive: Culture, Self-Confidence, Leadership;
negative: Sociability, Social Sensitivity, Tidiness). The most important deviation
from our hypotheses was for the Mature Personality scale, which was predicted to
have a negative quadratic association with g, but instead had a positive one.
4.3.1 LMS results compared to GAM results
Figure 4.3 shows a comparison of the fitted functions in the LMS and GAM
models for the example of Sociability in grade 10 males. As can be seen, the
predicted personality levels are similar in both models. In general, visual inspection
of GAM-predicted values showed a close correspondence with LMS results,
indicating that a combination of linear and quadratic effects gave a good
approximation of the relations revealed by the GAMs (other graphs of the GAMs are
available from the first author). In addition, the R2 for the GAMs were consistent
with the variance explained by the combination of the linear and quadratic effects of
LMS (slightly more variance was accounted for in the LMS models due to the use of
a latent g factor instead of factor scores). For the GAMs in males, the three
personality traits where g predicted the most variance were Sociability (3.3%),
Leadership (2.1%) and Tidiness (1.7%). In females it was Sociability (4.4%),
Mature Personality (3.7%), and Tidiness (2.5%). The variance explained in the
personality scales was higher for females than males in a number of cases, although
this varied greatly across scales (see Tables 4.4 and 4.5). Table B3 in Appendix B
compares the AICs for the GAM models to the null models in the grade 10 samples.
Compared with the null models, the GAM models all displayed lower AICs,
indicating that they added predictive power compared to a model with no predictors
(and were better fitting).
82
4.3.2 Grade and sex differences
The comparison of grade and sex differences in the estimated personality-
intelligence relations requires the assumption of measurement invariance between the
samples for intelligence and personality. This assumption could not be tested for the
personality scales due to lack of item-level data. In addition, measurement
invariance testing revealed that although configural and weak invariance were
tenable across grade and sex the VPR intelligence model, strong invariance (equality
of the intercepts) was not supported in both cases, as indicated by an decrease in CFI
> .010 (Cheung & Rensvold, 2002). Therefore, the differences in personality-
intelligence relations across samples must be interpreted with caution, as they may be
attributable to differences in the measurement of intelligence (or personality).
With this caveat in mind, there were some differences in the measured
relations across grade. Notably, the linear relation of Sociability with g was more
negative with increasing grade level (comparing grade 9 to grade 12 in males: ∆ beta
= -.88, log-likelihood ratio test: 2 (1) = 263.58, p < .001; in females: ∆ beta = -.118,
2 (1) = 514.98, p < .001). Two other important trends were the reduction of a
negative association of Leadership with g at higher grades (in males, ∆ beta =.106,
2 (1) = 387.06, p < .001; in females, ∆ beta = .080 2 (1) = 233.40, p < .001) and an
increase of the negative association of g with Tidiness (in males, ∆ beta = -.075, 2
(1) = 189.34, p < .001; in females, ∆ beta = -.099 2 (1) = 364.73, p < .001).
4.3.3 Figures 4.1 to 4.3: titles and captions
Figure 4.1 Mean personality as predicted by general intelligence (grade 10 males).
Caption: Personality scales and g are in standard units. Light lines represent 2
standard errors (SEs) above and below the mean (approximate 95% confidence
interval). SEs obtained from GAM models.
Figure 4.2 Mean personality as predicted by general intelligence (grade 10 females).
Caption: Personality scales and g are in standard units. Light lines represent 2
standard errors (SEs) above and below the mean (approximate 95% confidence
interval). SEs obtained from GAM models.
83
Figure 4.3 LMS and GAM-predicted sociability as a function of general intelligence
(grade 10 males).
Caption: LMS estimate = solid grey line. GAM estimate = dashed line. Sociability
and g are in standard units.
84
Figure 4.1
85
Figure 4.2
86
Figure 4.3
87
4.4 Discussion
In this study we examined linear and quadratic associations between g and
personality in Project TALENT. SEM was used to estimate linear and quadratic
effects of latent g on ten personality scales, and the influence of the general factor of
personality was controlled by residualizing the personality scores for the GFP. A
review of literature provided us with seventeen hypotheses of linear and quadratic
associations; nine of these hypotheses (53%) received support in all male samples
and ten (58%) received support in all female samples. In this section, we first review
the observed associations and discuss in greater detail some of the unexpected and
theoretically-relevant results. We then outline limitations of the study, and the
implications of our results for future research.
In divergence from previous studies (E. J. Austin et al., 1997; E. J. Austin et
al., 2002; Reeve et al., 2006) that have not done so, we found significant quadratic
associations of g with aspects of personality. Sociability, Vigor, Mature Personality
and Leadership were associated in this manner in males, and Sociability, Tidiness
and Mature Personality in females. These associations accounted for at most 3.9% of
the variance, so it would not be appropriate to conclude that prior studies have misled
the field in finding only small quadratic associations. Still, the associations we found
would have importance in considering mean personality scores of groups differing
greatly from average g. In our strongest example, using the grade 10 female sample,
negative linear and quadratic associations with g predicted a mean Sociability level
.70 SDs lower (SE = .02) for individuals two SDs above the mean on g, compared to
individuals of average g12
. Such a difference would generally be considered
substantive, though it did not render the mean Sociability level of 10th
grade females
with high g particularly unusual (the mean fell at approximately the 27th
percentile of
Sociability of the full sample). The group difference due to the linear effect alone
would be only .23 SD (SE = .02), accounting for 1.3% of the variance, compared to
3.7% of the variance in the model with the quadratic effect. This illustrates that
failing to consider nonlinear relations causes underestimation of the true associations
between certain personality traits and intelligence. Recognition of these nonlinear
12
For individuals with low g (two SDs below the mean), Sociability was .24 SDs below the mean.
88
associations can be particularly important when focus is on personality in groups
with extremely low or high levels of g and/or in understanding how the development
of personality and intelligence is intertwined in individuals.
Our results were generally consistent with previous findings on intelligence-
personality associations in samples of the general population and gifted adolescents
(Ackerman & Heggestad, 1997; Zeidner and Shani-Zinovich, 2011). We found that
males and females with higher g tended to have higher Self-Confidence, and males
with higher g also averaged higher Calmness; these scales reflected lower
Neuroticism in Five-Factor terms (Reeve et al., 2006). For the scales likely
reflecting Extraversion, Project TALENT participants with higher g scores tended to
display lower Sociability but higher Leadership, which was in line with the
hypothesis of Ackerman & Wolf (2005) that intelligence is linked to lower social
closeness but higher social potency. We found some indirect support for lower
Conscientiousness among more intelligent adolescents (for Tidiness, but not Mature
Personality), as observed by DeYoung (2011), and lower Agreeableness (Social
Sensitivity, but in girls only), as found in gifted studies (Sak, 2004; Zeidner and
Shani-Zinovich, 2011). Openness to Experience was incompletely represented by the
Culture and Leadership scales, but participants with higher g scores were above-
average on these scales, with the exception of Culture in grade 9 males. This
replicated the most common association in personality-intelligence studies
(Ackerman & Heggestad, 1997; DeYoung, 2011).
General intelligence was associated with mean-level differences in all Big
Five domains, which is somewhat at odds with existing theories of personality-
intelligence relations. For example, Chamorro-Premuzic and Furnham (2006)
maintain that each of the Big Five should be related to intellectual competence, but
regard Agreeableness as a marginal indicator, and view Neuroticism as mainly being
related to intelligence through test anxiety, and Extraversion related through test-
taking style. The opposing associations with g that we observed for Sociability and
Leadership (two aspects of Extraversion), cannot be well-explained within their
framework. PPIK theory also does not provide a full explanation for broad
associations between g and the Big Five. In PPIK theory, g is mainly associated with
89
personality due to the involvement of group abilities (such as crystallized intelligence
and perceptual speed) in particular trait complexes. Most notably, crystallized
intelligence is thought to contribute to the Intellectual/Cultural trait complex, along
with Openness to Experience (Ackerman & Heggestad, 1997; Ackerman & Beier,
2006). However, many of the associations we observed in the current study would
not be predicted in this framework, such as the association between higher g and
lower scores on scales reflecting Conscientiousness and Agreeableness, as well as
the differential associations of g with social closeness and social potency (Ackerman,
2005). For instance, Ackerman & Heggestad (1996) stated that: “Intelligence-as-
process correlates weakly with most broad personality factors, except [negatively]
for those that are associated with psychopathology” (p. 239).
Our results suggest instead that there are meaningful associations between g
and each of the Big Five (and/or their facets). Moreover, g is closely related to
intelligence-as-process or fluid intelligence (Gustafsson, 2002; Kvist & Gustafsson,
2008; Major, Johnson, & Deary, 2012). Overall, our results were more consistent
with the personality differences observed in studies of gifted adolescents, such as a
greater tendency towards Perceiving (which is correlated with lower
Conscientiousness; McCrae & Costa, 1989) and Introversion on the Myers-Briggs
Type Indicator (Sak, 2004) and lower Agreeableness in the Big Five (Zeidner and
Shani-Zinovich, 2011). Because personality differences have been more apparent in
these studies, it seems likely that considering the developmental differences between
gifted and normally-developing children and adolescents may be a good way to
develop theories of personality-intelligence associations, in addition to examining
associations in the general population. It is possible, however, that gifted people
may have distinct life experiences (such as experience with accelerated education
programs) that make comparisons with general samples more difficult.
Due to quadratic associations of g with personality, adolescents with low g
did not necessarily display the converses of the personality associations of those with
high g, and in fact were more similar in score with high-g students than average
ability students on a number of scales. For example, like high-g students, they
averaged lower Sociability and Tidiness. Participants with low g were also
90
unexpectedly found to average higher scores on the Mature Personality and
Leadership scales than average-ability students.
Due to the positive association between Mature Personality and
Conscientiousness in Reeve et al. (2006), and the negative association of g with
Conscientiousness in the literature, we predicted a negative quadratic effect of g on
Maturity. The unexpected positive quadratic effect may reflect the fact that the
Mature Personality scale contained several items that tapped self-assessed
achievement striving and engagement (“I work fast and get a lot done”; “I am
productive”). Thus, it may not be so surprising that students with higher g scores
(who also tended to have had more success at school) also obtained higher scores on
this scale, possibly despite its association with Conscientiousness, on which they
tended to score lower.
The most unexpected linear association was the negative association between
g and the Leadership scale, mostly found in the lower grade levels (in grade 9
males/females, beta = -.12/-.11). This finding may have reflected lack of clear
understanding of the items by younger and less able students (e.g. the item “I am
influential”). Another possibility is that the students understood the items, but that
less intelligent students overestimated their leadership abilities due to lower
metacognitive ability to assess their social function (Kruger & Dunning, 1999). This
‘Dunning-Kruger’ effect may also apply to our finding of higher scores on the
Mature Personality scale for individuals with below-average g. One final possible
interpretation for the Leadership scale finding is that leadership in younger grades is
often more social than intellectual in nature, and that social engagement may be
negatively related to intellectual performance due to an investment trade-off between
social and intellectual activities (Ackerman & Wolf, 2005).
Most of the linear and quadratic associations we observed were present in
both sexes. The exceptions to this were linear associations of g with Culture,
Impulsiveness and Calmness, and a quadratic effect on Social Sensitivity that was
only present in females. Based on the content of the Culture scale and its positive
association with Big Five Openness (Reeve et al. (2006), we predicted that Culture
would show a positive association with g. This was the case in females, but small
91
negative associations were found in males (beta = - .067 to -.087). This may reflect
the fact that the Culture scale emphasized having good manners over intellectual
interests. Perhaps the socialization pressures on girls to be well-mannered were
stronger than those on boys. The other sex differences we observed were less readily
interpretable.
For scales where an association was found at any grade level, the majority
were found in all four grade samples: 13 of 19 in males (68.4%) and 14 of 18 in
females (77.8%). This observation supports the view that the effects were not due to
chance measurement artifacts from individual samples.
As noted in the Results, even if most effects were consistent, some effects
varied substantively in magnitude across grades. These differences have also been
the subject of prior theories on personality-intelligence relations (Ackerman & Wolf,
2005; Chamorro-Premuzic & Furnham, 2006). The increase of the negative
quadratic effect of g on Sociability across grades provides support for the hypothesis
that higher social closeness may run counter to the development of intelligence
because adolescents with greater need for social closeness select social activities
more frequently than (generally) solitary intellectual activities (Ackerman & Wolf,
2005). However, the direction of effect assumed and measured in this study may
imply that it was instead higher intelligence that reduced social closeness over time
due as higher-g students increasingly selecting more solitary activities. Due to the
unavailability of item-level data, the PT personality scales were not well-suited to
testing the quadratic effect of personality on intelligence (the effect hypothesized by
Ackerman & Wolf, 2005). Future studies may be able to disentangle these two
effects by comparing the sizes of quadratic effects in each direction.
The negative linear association of g with Tidiness became stronger with grade
level, which is consistent with the hypothesis of Chamorro-Premuzic and Furnham
(2006) that lower intelligence leads to the development of greater orderliness over
time as a compensatory mechanism to meet environmental demands13
. However, the
13
We tested this hypothesis alternatively by examining differences in mean Tidiness with the sample
split by quintile of g. Comparing grade 9 to grade 12 samples, Tidiness increased significantly in
every quintile (p <.001), but there was a progressively greater increase in Tidiness corresponding to
lower quintile of g.
92
negative quadratic effect of g on Tidiness also indicated that very low levels of g
corresponded with decreased Tidiness, possibly because very low intelligence is a
hindrance to orderly behavior.
There are several limitations surrounding our conclusions regarding
personality-intelligence relations. First, our personality scales may not have been
measurement invariant across different levels of g, which could have caused apparent
linear and nonlinear associations that did not exist (McLarnon & Carswell, 2012;
Waiyavutti, Johnson, & Deary, 2012). We had, however, no way to test this as we
did not have access to the items. Second, we were able to establish that measurement
invariance did not hold across samples for g. Thus, although we observed some
consistency of associations across grades and sexes, the constructs measured across
the samples may not have been identical.
The PT personality data had a large GFP that accounted for approximately
40% of the variance in each sample. Our removal of the GFP may have been a
limitation because it may have contained substantive personality variance, although
most recent research supports a largely artifactual origin of the GFP (Anusic et al.,
2009; M.C. Ashton et al., 2009; Bäckström et al., 2009; Chang et al., 2012). One
possible explanation for the large GFP in Project TALENT is that the context of in-
school testing may have influenced students to “fake good” on the personality scales,
and the more intelligent students were more capable and/or more motivated to do so.
Given that the main purpose of PT (of which the students were aware) was to assess
scholastic talent, it would be most relevant for students to exaggerate scores on
scales tapping behaviour socially desirable in the school context (such as diligence
and responsibility). The high loading of the Mature Personality scale on the GFP
(mean r = .79) was consistent with this interpretation, as was the relatively higher
loading of Tidiness on the GFP in PT samples compared with the college sample of
Reeve et al. (2006). A possible non-artifactual explanation is that more intelligent
students were in fact more successfully socialized within the high school
environment, and that this led to higher scores on all the PT personality scales.
Discounting this, however, studies of students selected for high intelligence found
that they did not score higher than unselected students on Agreeableness and
93
Conscientiousness—Big Five factors that reflect greater socialization (Sak, 2004;
Zeidner & Shani-Zinovich, 2011). The alternative of including the GFP could have
led to exaggerated g-personality associations.
Controlling the GFP in the current study caused a number of the personality
scales to have negative linear associations with g, in contrast with the results of
Reeve et al. (2006), who found only positive linear associations. Nonetheless, out of
eight positive-direction associations in Reeve et al. (2006), six were also found here.
The exceptions were the positive associations of g with Social Sensitivity and
Calmness in females. Replicated associations were with Mature Personality,
Calmness and Self-Confidence in males; Mature Personality, Culture, and Self-
Confidence in females. These results confirm that the GFP was not entirely
responsible for the positive associations observed in the previous study.
Nonetheless, one key implication of our results is that the GFP can be a
potentially important confounder or mediator of personality-g associations,
particularly for linear associations as these relations were the most affected by
removal of the GFP (see Supplemental Tables B1 and B2). If the GFP represented at
least partly substantive variance instead of methodological variance, it could have
been a mediator, whereby the effect of g on personality occurred indirectly through
the GFP, or vice-versa.
One final limitation to our study was that the sample was assessed in 1960,
and relations between personality and intelligence may have shifted since then. This
kind of change was observed by Wolf and Ackerman (2005), who that the relation
between Extraversion and intelligence was slightly positive before 2000, but slightly
negative after 2000. One notable source of such change concerns the erosion of
gendered occupational roles since then. Girls at that time had less opportunity to
aspire to high education, and especially to occupational achievement in their own
names. Moreover, it was very common that they aspired and expected to marry and
be supported financially by their husbands. Despite this possibility, females had
higher scores than males on the Mature Personality scale at all grade levels, and the
association between g and Maturity was higher in females than males (see Tables 4.4
and 4.5). Offsetting the age of the sample, one of the strengths of the Project
94
TALENT sample was that it was representative of the United States in 1960
(Flanagan et al., 1962), so that our results can be generalized to the whole population
at that time.
4.4.1 Conclusions and future directions
We found that mean levels for most Project TALENT personality scale scores
varied substantially across levels of g, and a number of scales showed quadratic
associations. These results provide further support for the view that personality-
intelligence associations are substantive and relevant to understanding the
development of individual differences in both domains (Ackerman, 1996; DeYoung,
2011; Sak, 2004). Our results also indicated two directions for future research in this
area: the interpretation of the general factor of personality, and the use of nonlinear
models to test the direction of effect (personality on intelligence, or intelligence on
personality).
If it was not controlled, the GFP would have had a substantial effect on
personality-intelligence relations in the current study due to its positive association
with g. Future research should examine whether this relation is substantive or
artifactual in nature, possibly through the use of multiple raters or social-desirability
scales. If the GFP itself is found to be largely artifactual, as much recent research
suggests, then it is questionable whether the g-GFP association can represent
meaningful variance, but research in this area is still ongoing.
The potential to examine direction of effect deserves more consideration in
personality-intelligence research. In the current study, we focused on the
associations of the quadratic function of g with personality, but such nonlinear
associations may be found in the other direction, or in both directions. Although the
nonlinear associations we observed were small in terms of variance explained, they
were capable of resulting in substantive personality differences for individuals at the
extreme ends of the g distribution. Nonlinear associations that result in substantive
differences in personality at the tails of the intelligence distribution, or differences in
intelligence at the tails of personality distribution, can potentially be very informative
about how personality and intelligence interact with each other. In spite of this, it is
likely that the direction of influence runs both ways in most cases, and that the
95
strength and direction varies depending on the environment and over time. In order
to understand the interplay between personality and intelligence more complex study
designs and models are needed.
4.5 Integrating cognitive abilities and interests
After examining personality-intelligence associations in PT, the next step was
to bring occupational interests into the research. As presented in the introduction, I
decided to focus on the possibility of replicating the trait complexes composed of
cognitive abilities and interests in PPIK theory (Ackerman & Heggestad, 1997). The
proposed capacity of these trait complexes to predict future occupation is a key
aspect of the theory, and a strength of the PT dataset was its longitudinal data. Thus
the trait complexes were examined in the context of occupation eleven years after
high school, where the hypothesis was that they should have equal predictive validity
to the use of individual scores for cognitive abilities and interests. This was tested
for trait complexes composed both of factors and latent classes. In addition, the trait
complexes were evaluated in terms of how well they matched their descriptions in
PPIK theory.
96
Chapter 5: Trait complexes of cognitive abilities and interests and their predictive validity for occupation
5.1 Introduction
Intelligence and interests are both important predictors of occupational
attainment and job type (De Fruyt & Mervielde, 1999; Kuncel, Hezlett, & Ones,
2004; Schmidt & Hunter, 2004). General intelligence (g) relates strongly to
occupational level (for review, see Schmidt & Hunter, 2004) Specific abilities such
as spatial and verbal abilities are also relevant to employment in specific
occupational areas such as the humanities and scientific fields (Johnson & Bouchard,
2009; Wai, Lubinski, & Benbow, 2009). In keeping with their applied purpose,
measures of occupational interests are predictive of the nature of future employment
(J. T. Austin & Hanisch, 1990; De Fruyt & Mervielde, 1999), as well as of
performance in that employment (Van Iddekinge, Putka, & Campbell, 2011).
Given the predictive validity of interests and cognitive abilities, discovering
any overlap between them could have theoretical as well as applied significance
(Johnson & Bouchard, 2009). Greater understanding of the links between cognitive
abilities and interests could aid our theories of both cognitive and interest
development. Some researchers have proposed that cognitive abilities have
substantial roles in the development of interests. Gottfredson’s (1986, 2005) theory
of circumscription and compromise posits that individuals’ self-awareness of their
levels of general intelligence influences their interest in particular occupations
according to their cognitive complexity. Hogan and Roberts’ (2000) socioanalytic
model of identity development proposed that interests are built on successful
experiences with cognitive investment, which depend on intelligence. In turn,
interests are theorized to influence the development of intelligence through the
selection of future learning environments (Hogan & Roberts, 2000; Scarr, 1996).
Development of a framework that better integrates interests and cognitive abilities
could lead to better career counseling advice, and long-term increases in person-
environment fit, which refers to matches between work environments and
individuals’ abilities and preferences (Dawis & Lofquist, 1984; Holland, 1997).
97
Increases in person-environment fit could have benefits for both individuals’ work
satisfaction and productivity.
In spite of the potential importance of integrative research, there have been
relatively few studies of the associations between interests and cognitive abilities
(Ackerman & Heggestad, 1997; Anthoney & Armstrong, 2010; Johnson &
Bouchard, 2009). Moreover, these studies have often been hindered by the use of
college samples with restricted ability ranges, and failure to separate specific abilities
from general intelligence statistically and conceptually (Johnson & Bouchard, 2009).
For example, Ackerman & Heggestad (1997) reported meta-analytic associations of
interests with nominally specific abilities, but these ability measures were not
statistically independent of g, leaving it unclear to what degree the associations were
ascribable to g or specific abilities. At the same time, general theories of interest-
ability interaction such as Hogan’s and Gottfredson’s have not provided detailed
hypotheses on the overlap between specific abilities and interests, predictions which
might be the most useful for practical and theoretical reasons. In spite of the
limitations of Ackerman and Heggestad’s (1997) meta-analysis, PPIK theory
(intelligence-as-process, personality, interests, and intelligence-as-knowledge) by
Ackerman (1996) did address this specific overlap.
Ackerman and colleagues proposed that intelligence, personality and interests
coalesce into four “trait complexes” (Social, Clerical/Conventional, Science/Math,
and Intellectual/Cultural) (Ackerman, 1996; Ackerman & Heggestad, 1997). They
defined trait complexes as similar to Snow’s (1963) aptitude complexes, or
“combinations of levels of some variables which are particularly appropriate for
efficient learning” (p. 120, cited in Ackerman & Beier, 2003a); however, Ackerman
and colleagues focused on attainment of academic knowledge/expertise and practice
of particular occupations rather than the learning processes necessary to reach those
states. Like Hogan & Roberts (2000), they regarded the interaction between interests
and cognitive ability as reciprocal. They proposed that particular abilities and
interests become more strongly related throughout development because the abilities
are suited to success in certain domains, and the satisfaction brought by this success
spurs the interest that motivates further cognitive investment. Trait complexes are
98
thus held to influence subsequently how individuals select and attain particular
occupations through their roles in knowledge acquisition in academic/occupational
settings (Ackerman, 1996).
In PPIK theory, interests are conceptualized in terms of the RIASEC model
of occupational types by Holland (1973, 1997; Realistic, Investigative, Artistic,
Social, Enterprising and Conventional). The RIASEC model is the predominant
model of occupational interests in the literature. It is based on preferences for six
types of work environments that are organized in a hexagonal circumplex, with
adjacent types more closely related in job demands than opposite types. Research on
the RIASEC hexagon has supported this structure for interests and employment types
in the United States (Holland, 1997; Tracey & Rounds, 1993) A significant
limitation is that the model does not address the roles of cognitive abilities in
interests. However, given the consistent associations between RIASEC interests and
cognitive ability measures that Ackerman and Heggestad (1997) found, they used its
framework to formulate PPIK theory (Ackerman, 1996). The RIASEC model is
well-suited to an integrative framework because work environments are important
(though not the only) contexts in which ability and non-ability traits converge
(Armstrong et al., 2008).
PPIK theory also specified that only two of the four trait complexes primarily
involve ability-interest associations; the other two primarily involve personality-
interest associations (Ackerman, 1997). Based on a meta-analysis of five studies,
Ackerman and Heggestad (1997) observed that only three of the six Holland types
were related substantially to cognitive abilities: Realistic interests, defined as
interests in activities involving physical action and motor coordination; Investigative
interests, defined as interests in cognitive problem solving; and Artistic interests,
defined as interests in expression through artistic media (Holland, 1973). Realistic
and Investigative interests were associated, and each was also associated with
general intelligence (intelligence-as-process in PPIK theory), as well as with math
and spatial abilities, relations which formed a Science/Math trait complex (for
example, the meta-analytic correlation of spatial ability with Realistic interests was
.28; Figure 5; Ackerman, 1996). Artistic interests were held to be associated with
99
crystallized or verbal ability (intelligence-as-knowledge in PPIK theory), forming an
Intellectual/Cultural trait complex (e.g. a correlation of .36 between Artistic interests
and verbal ability; Ackerman, 1996). The other two trait complexes (Social and
Clerical/Conventional) showed moderate associations with personality, but only
minor and/or less consistent associations with cognitive abilities.14
While personality
is likely to be important in understanding occupational outcomes in certain domains,
it may be helpful to further the understanding of interest-ability associations
exclusive of personality before attempting to integrate all three sources of individual
differences. The latter was the focus of the current study. As we did not include
personality in our analyses, we anticipated finding only the Science/Math and
Intellectual/Cultural trait complexes.
Ackerman and colleagues found that their trait complexes were moderately to
strongly related to academic knowledge (Ackerman & Rolfus, 1999), selected
university course (Ackerman, 2000), and university course performance (Kanfer et
al., 2010), which are outcomes along the path to vocational choice. However, trait
complexes in these studies were obtained through factor analysis, which may be
inconsistent with how trait complexes have been conceptualized. Factor analysis is
used to group variables, and relies on the assumption that the groupings apply in the
same way to all individuals in the population. A method of analysis that is arguably
better suited to identifying trait complexes is latent class analysis (LCA). LCA
groups individuals together based on their scores of a set of variables, ignoring the
associations among the variables at the population level. Thus, latent classes can
represent groups of individuals who have “combinations of levels of some variables”
(Snow, 1963), rather than sets of positions on groups of variables.
One study that used LCA to define interests groups was conducted by
Johnson and Bouchard (2009). They examined the mean-level differences in
cognitive abilities among eight latent interest classes, where cognitive ability was
defined according to an updated version of Vernon’s intelligence model, the Verbal,
14
The Conventional trait complex involved a positive correlation between Conventional interests and
perceptual speed (e.g. .15 in Rolfus & Ackerman, 1996), and the Social trait complex small negative
correlations between Social interests with math ability (-.21 in Rolfus & Ackerman, 1996) and spatial
ability (-.13 in Randahl, 1991).
100
Perceptual and Image Rotation (VPR) model (Johnson & Bouchard, 2005b). Based
on previous research in the same sample, Johnson & Bouchard (2009) separated
intelligence into orthogonal factors for general intelligence (g) and two residual
dimensions: Verbal-Image Rotation and Focus-Diffusion (Johnson, Bouchard, et al.,
2007). Mean levels of g varied strongly among the eight latent classes. Beyond this,
however, latent classes of interests in leadership, exploration and adventure were
related to Image Rotation abilities, whereas interests in cultural and persuasion
occupations were related to Verbal abilities. These results were consistent with
Vernon’s (1961) conceptualization of interests in his verbal-perceptual model of
intelligence. Vernon proposed that verbal and math abilities were related to
achievement and interest in traditional educational (math and verbal subjects). On
the other hand, perceptual (spatial and mechanical) abilities were related to aptitude
for technical, scientific and practical subjects. Although Ackerman and Heggestad
(1997) did not statistically isolate g from specific abilities, these two themes were
apparent in their Math/Science and Intellectual/Cultural trait complexes. In addition,
Johnson and Bouchard’s (2009) results supported the important role of g in
occupational interests (Gottfredson, 1986, 2005), such that mean levels of
intelligence varied strongly across the interest classes, being highest for Science and
lowest for Personal Care. This role for g is not as central in the trait complexes of
PPIK theory.
Johnson and Bouchard’s (2009) study was not intended to test the concept of
trait complexes. In order to examine whether there are trait complexes of interests
and abilities, both variables should be entered simultaneously into latent class
analysis, so that groups with different levels of interests and abilities may be
identified. These groups would then reflect the integration of interests and abilities
in interlocked transactions implied by Ackerman and Beier’s (2003) definition of
trait “trait complex”. The main purpose of the current study was to test the validity
of the trait complex concept. This was done by comparing the ability of trait
complexes to predict occupational type with the individual scale scores for cognitive
ability and interests. We hypothesized that if trait complexes are true groupings of
individuals which influence the likelihood of acquiring specialized occupational
knowledge, and thus influence career choices (Ackerman & Beier, 2003a), then
101
latent classes representing them should show strong predictive validity for type of
future employment. Ackerman and Beier (2003) put this question similarly: “is there
a synergy among elements within the trait complexes, so that concentrating on trait
complexes is more informative in the career choice context than individual trait
measures?” (p. 209). Due to the prominent position that trait complexes occupy in
PPIK theory, we predicted that they should demonstrate predictive power (in terms
of explained variance) at least equal to the use of individual scores for cognitive
abilities and interests. Equality would be accepted as being in favour of trait
complexes because of the greater parsimony of latent classes. As a point of
comparison with the research of Ackerman and colleagues, we also compared the
predictive validity of factor-analytic trait complexes to the individual scale scores,
with the same predictions for these trait complexes as for the LCA trait complexes.
The current study was thus meant to address two questions: first, would
exploratorily-derived trait complexes of interests and abilities, obtained through LCA
and factor analysis, replicate the content of the Science/Math and
Intellectual/Cultural trait complexes proposed by Ackerman and Heggestad (1997)?
The Social and Clerical/Conventional trait complexes were not anticipated due to the
exclusion of personality variables in the present analysis. Second, would the trait
complexes obtained by either method display predictive validities at least equal to
those of individual scale scores? The predictions of PPIK theory would be
contradicted if the trait complexes obtained did not resemble the Science/Math and
Intellectual/Cultural trait complexes, and/or if the trait complexes did not display
predictive validities comparable to those of individual scale scores. Additionally, if
only the trait complexes derived by factor analysis satisfied these conditions then it
would raise theoretical questions about their definition as combinations of levels of
the variables.
5.1.2 Previous Project TALENT research
For this study, we made use of data from Project TALENT (PT). PT was a
longitudinal study of American high school students meant to investigate their
aptitudes, interests, and backgrounds, and the influences of these variables on
educational and occupational outcomes (Flanagan et al., 1962). During the study, 60
102
aptitude tests and extensive interest scales were given to a sample of over 400,000
American high school students, who were representative of the U.S. student
population. The students were followed up over an 11-year period after high school
and surveyed on their education and occupational experiences.
Two previous have studies have used PT intelligence and interest data to
predict occupational type (Austin & Hanisch, 1990; Humphreys, Lubinski & Yao,
1993). Austin and Hanisch (1990) examined the tenth-grade PT sample and found
five discriminant functions that predicted the 12 occupation categories defined by PT
investigators. The results indicated that occupational category could be predicted
above chance for 10 of the 12 categories (exceptions were Technical and Sales jobs).
Two major discriminant functions described the interest and ability data. The first
discriminant function, interpreted as verbally-oriented general mental ability, mainly
predicted occupational prestige or level. The second function, which differentiated
individuals based on mathematics, spatial ability and gender, predicted scientific and
technical occupations (Hanisch & Austin, 1990). These functions, however, were
not interpreted in a trait-complex framework, and their capacities to predict
occupation were not compared to those of individual scales.
Humphreys et al. (1993) explored the differential prediction of occupation for
groups defined by the top 20 percentiles of spatial and verbal abilities. They
attempted to equalize these groups for general intelligence by selecting students in
the top 20 percentiles on composites of spatial-math and verbal-math scores, math
ability being used as a proxy for general cognitive ability. They found that the high-
verbal and high-spatial groups had significantly different probabilities of entering
scientific/engineering and humanities jobs. The groups also differed strongly in their
mean occupational interests, with the high spatial-ability group showing greater
interest in mechanical-technical jobs, and the high verbal-ability group more
interested in literary-linguistic jobs. Although these groups could be considered
similar to trait complex groups, they were pre-specified and not derived through any
empirical analysis. In addition, the composite method of Humphreys et al. (1993)
was not entirely successful in controlling for g, as the high verbal group had higher
mean scores for math than the high spatial group (see Table 3 of Humphreys et al.,
103
1993). As in other studies on ability-interest associations, a significant limitation of
both Austin and Hanisch (1990) and Humphreys et al. (1993) was the lack of
independence of the specific ability measures from general intelligence. In
summary, the current study tested the trait complexes of PPIK theory in two novel
ways, by employing latent class analysis to define trait complexes, and by comparing
their predictive validity for future occupation with individual scale scores for
cognitive abilities and interests.
5.2 Method
5.2.1 Sample
The two highest-grade samples of PT were used (grades 11 and 12), because we
considered the older students more likely to have considered their future career
prospects. The use of two samples allowed for replication of potential trait
complexes across samples. In grade 11, there were 47,027 females and 45,292
males. In grade 12, there were 41,456 females and 39,674 males. The total sample
size was 173,449, with 51.0% females. The mean age was 16.4 in grade 11 (SD =
.69) and 17.3 in grade 12 (SD = .67). Males were slightly older than females in both
samples: 16.4 compared to 16.3 in grade 11, and 17.4 compared to 17.2 in grade 12.
The full individual age range was 8 to 21, the younger participants having skipped
multiple grades, and the older participants having been held back.
For the 11-year follow-up mail survey, responses were obtained from 27.5% of
the original grade-11 sample, and 30.9% of the grade-12 sample. In order to adjust
for the lack of representativeness of the follow-up sample, PT investigators
conducted special interviews with non-respondents to the mail questionnaires.
Approximately 2500 participants in each grade cohort were given telephone or in-
person interviews (Wise et al., 1979). Sample weights were created in accordance
with the sampling ratio of the special sample to original the 1960 sample (Wise et al.,
1979). We applied these sample weights to our analyses where follow-up occupation
data were used. The follow-ups were conducted in 1971 for the grade-12 sample and
1972 for the grade-11 sample. Participants were not asked their ages at follow-up;
however the dates of the follow-up surveys were recorded. The mean week of the
year that the surveys were received was week 7.1 in the grade-12 sample (SD = 6.3)
104
and week 8.2 for the grade-11 sample (SD = 8.0). As the baseline data were
collected in March 1960, it can be inferred that the mean follow-up sample age was
approximately ten years and eleven months older than baseline for the grade-12
sample. For the grade-11 sample it was eleven years and eleven months older than at
baseline. Thus, their mean ages were approximately 28.3 for the grade-11 sample
and 28.2 for the grade-12 sample. The follow-up sample was 52% female according
to gender recorded at baseline.
5.2.2 Intelligence measures
The intelligence measures were chosen from the 60 aptitude and achievement
tests in PT (Wise et al., 1979). To ensure that our measures of specific ability did not
rely on overly specialized knowledge, we excluded the “information tests” on
academic and non-academic topics. Our starting point was the 22-test narrow
selection as defined in our previous study (Major et al., 2012; see descriptions of the
tests and reliabilities there). However, we excluded three of the English achievement
tests (Spelling, Capitalization and Punctuation) because they relied too heavily on
knowledge acquired in school classes. The Vocabulary test was also excluded
because it was part of the original information tests. Eighteen tests remained after
this selection. In previous studies, the advanced math test was omitted because it
was designed for students above the tenth grade, and was thus deemed unfair for
younger students. However, it was also excluded here because we initially planned
to use all four grade samples.
Data screening was the same as in Major et al. (2012). Scores on the PT
response credibility index, based on a screening test assessing illiteracy, mental
disability or apathetic testing attitude, were used to exclude participants who did not
reach the cut-offs set by the PT study designers, except where only mental slowness
was indicated (a low score for the number of responses on the Clerical Checking
test). Transformations were applied to two tests that displayed non-normal
distributions (English Usage and Table Reading): English Usage was negatively
skewed, and we applied a square-root transformation (the direction of the variable
was reversed prior to transformation and then reversed back). Table Reading was
strongly positively skewed and leptokurtic and thus had logarithmic and cosine
105
transformations applied. Cases showing severe problems with multivariate outliers
were also removed (Major et al., 2012). Following data screening of the intelligence
tests, total sample size was reduced to 170,723 (1.57% of the total sample removed,
the majority for failure on the response credibility index). After removal of these
cases, some missing data remained for each cognitive ability test. In the male
samples, 2.1 to 2.8% of scores were missing, while 2.1 to 2.7% were missing in the
female samples.
5.2.3 Interest measures
The PT interest scales consist of 17 composites that were designed to capture
interests in different job areas, such as Artistic and Mechanical-Technical jobs (Wise
et al., 1979). However, a limitation of these scales is that they were created on an a
priori basis, without regard to the observed correlations among the items, or to any
particular theoretical framework. Because item-level data were available, we derived
new interest scales based on exploratory factor analysis. Previous studies have
employed the pre-existing scales (J. T. Austin & Hanisch, 1990; Humphreys,
Lubinski, & Yao, 1993). It was anticipated that these new scales would have greater
validity and thus predictive power for occupation than the original PT interest scales
The original interest scales were formed from 205 items, 122 of which were
occupation titles (e.g. musician, rancher, etc.) and 83 of which were activities
applicable to work and school settings (e.g. typewriting, selling furniture). Students
were asked to indicate how well they would like or dislike the occupation or activity,
and were instructed to disregard educational requirements, salary, social standing, or
other factors (Wise et al., 1979). Responses were recorded on a 5-point Likert scale
from “I would dislike this very much” to “I would like this very much”.
For the occupation titles, missing data percentages for the items for each
sample ranged from 2.6-6.7%, and were very consistent across grade and sex. An
exception was 21 consecutive items in each sample that had greater numbers of
omitted responses, likely due to a coding error in the PT database. For example, in
grade-12 males, these 21 items had total missing data percentages of 7.9-9.0%.
Because these additional missing data were missing at random, they should not have
biased the analyses. Three occupation items were excluded from the male samples
106
because they were female-oriented occupations that received very low endorsements
(Maid, Dish Washer and Housewife).
5.2.4 Occupational categories
Table 5.1 contains the occupation category titles and sample percentages for
grade-12 males and females with 11-year follow-up data. Participants were asked to
state their current job titles. These written responses were reduced to 254 job codes
representing specific jobs or job areas such as Airplane Navigator, Veterinarian or
Metal Trades (Wise et al., 1979). These job titles were then organized by PT
investigators into twelve categories according to broad occupational themes. The
most prevalent job category in males was Business Administration, while in females
it was Clerical and Office Work. Due to the period during which the data were
collected, there were large gender differences in the frequency of different
occupational groups.
Table 5.1 Occupation categories and sample percentages (grade 12 sample).
Category title Males (%) Females (%)
Physical Sciences, Engineering and Mathematics 5.5 0.2
Medical and Biological Science 2.5 3.1
Business Administration 19.3 3.1
Teaching and Social Service 8.1 8.3
Humanities, Law, and Social Science 3.4 0.9
Fine and Performing Arts 1.0 0.4
Technical 5.3 1.9
Sales 11.2 1.9
Mechanical and Industrial Trades 8.9 0.5
Construction 7.8 0.02
Clerical and Office Work 3.4 14.7
General Labour and Public Service 15.2 7.6
Vague and Undesignated 8.3 7.9
Housewife N/A 49.5
5.2.5 Method of analysis
The analysis was done in three steps. First, factor analysis was performed
separately on the interest items and cognitive ability scales, and the factor scores
were retained. Second, trait complexes were obtained from the interest and cognitive
ability scores through both factor analysis and latent class analysis, conducted in
Mplus 5.21. In the third step, the trait complex data (class membership and factor
107
scores) were used to predict future occupational category using logistic regression in
Mplus and multinomial regression in SPSS 18. As the first step was preliminary to
the two main analyses, its results are presented in this section.
Missing data were handled through maximum likelihood estimation, which
assumes that the data were missing at random (MAR). This assumption was tenable
because it is unlikely that students purposely avoided particular cognitive ability or
interest tests.
5.2.6 Interest and cognitive ability factors
The 205 items were allocated to 17 original interest scales by PT designers on
an atheoretical basis. We derived new scales based on exploratory factor analysis
(EFA). Two separate EFAs were conducted for the occupation titles and activities.
The analysis was conducted in SPSS with Promax rotation (kappa = 3).
Examining the scree plots in the grade-11 and -12 male samples suggested
seven factors, while it was less clear in the female samples. When further factors
were extracted beyond seven in males, these also displayed adequate simple structure
and interpretability up to the tenth factor. Upon extracting ten factors in females,
nine of these were recognizable counterparts to the male factors, except the tenth,
which was not easily interpretable and obtained no loadings above .35 in either the
grades-11 or -12 samples. Thus, ten interest factors were retained in both males and
females.
The names assigned to the factors in grade-12 males and their two highest
factor loadings were as follows: Trades (Riveter: .77, Bricklayer: .73), Politics (U.S.
Congressman: .98, U.S. Senator: .96), Science (Chemical Engineer: .78, Electrical
Engineer: .76), Business Clerical (Bookkeeper: .75, Office Clerk: .69), Arts (Artist:
.87, Writer: .73), Military (Air Force Officer: .77, Army Officer: .74), Teaching
(High School Teacher: .92, School Principal: ..78), Medical (Doctor: .80, Surgeon:
.79), Business Sales (Stock Salesman: .55, Insurance Agent: .55), and Architecture
(Designer: .49, Architect: .48). The total variance explained by the ten factors was
45.1%.
108
The factors in grade-12 females were assigned the same names as the males;
the following were the two highest loadings in the grade-12 sample: Trades (House
Painter: .72, Deliveryman: .70), Politics (U.S. Congressman: .96, U.S. Senator: .93)
Science (Electrical Engineer: .74, Chemical Engineer: .67), Business Clerical
(Typist: .77, Secretary: .75), Military (Air Force Officer: .83, Marine Corps Officer:
.80), Teaching (High School Teacher: .70, School Principal: .60), Medical (Nurse:
.70, Doctor: .70), Business Sales (Insurance Agent: .55, Personnel Administrator:
.55). The total variance explained by the nine factors was 41.8%.
Factor analysis was also performed on the 83 activity items. Judging by scree
plots and interpretability, six factors were found in the grade-12 male sample.
However, upon examining their contents and correlations with the occupation title
factors, it was found that five out of six of these factors were redundant with the
occupation title factors. For example, there was a factor composed of activities
relevant to trades occupations (e.g. “repair an auto”, “work in a steel mill”), which
correlated highly with the trades factor for the occupation titles (r = 0.76). Activities
factors were found with moderate to high correlations with the occupation factors
that were labelled Science, Business Clerical, Arts, and Business Sales (mean
correlation = .65, SD = .12). The remaining activity factor was composed of Sports
activities, and correlated most highly with the Military occupation factor (r = .35).
Because of the greater robustness of the occupation title factors (due to the greater
number of items), and the redundancy of the activities factors, we decided to retain
only the occupation title factors to represent the occupational interests. The activity
factors were not used.
Due to the presence of missing data, the factor scores for the interest factors
could not be obtained without the exclusion of participants with incomplete data.
Thus, we constructed composites using the uniformly-weighted means of the non-
missing item scores, selecting items that loaded .30 or above on the respective
factors. The composites correlated highly with the factor scores for participants with
complete data. Excluding the Architecture composite, in grade-12 males, the
composites had correlations with their factor scores that ranged from .92 to .99. The
correlation of Architecture composite with its factor was lower (.53). This was due to
109
the lower factor loadings for the Architecture factor (see above). In grade-12
females, the composites had correlations with the factor scores that ranged from .93
to .99. The range of correlations was similar in the grade-11 samples: in grade-11
males it was .78 to .99 (Architecture: .45), in grade-11 females it was .94 to .99.
As in Johnson & Bouchard (2007), we sought to obtain specific cognitive
ability scores that were separate from general intelligence. To do this, we extracted
the general factor of the 18 tests using maximum likelihood estimation and obtained
g factor scores. The g factor explained a mean of 35.6% of the variance in the male
samples, and 36.4% in the female samples. We regressed the individual test scores
onto the g-factor scores, and entered the residuals into EFA in Mplus. In all four
samples, the scree plots suggested four residual factors and the 4-factor EFA solution
displayed good fit statistics. For example, in grade-12 males these were: RMSEA,
.043 (90% confidence interval: .042 - .044), SRMR, .026; in grade-12 females:
RMSEA: .042 (.041- .043), SRMR: .018. The four residual factors were labelled
Spatial (made up of tests requiring spatial reasoning), English (loadings from the
English tests and Memory for Words), Speed (loadings from all the speeded tests),
and Math (a bipolar factor on which the Math tests and Arithmetic loaded positively,
and three Verbal tests loaded negatively).
Based on these exploratory results, we created a confirmatory bi-factor model
in each sample. In the bi-factor model, g is allowed to influence each test score, and
specific abilities form their own factors that are uncorrelated with g. Factor loadings
on the four factors were specified if they were .15 or greater in the EFA results.
Table 2 displays the factor loadings for the confirmatory bi-factor models in the
grade-12 samples. In the grade-12 males the specific ability factors accounted for
2.0% to 9.1% of the variance, in the grade-12 females they accounted for 1.5% to
8.5% of the variance. The same factor model forms were specified in the grade-11
samples, but we allowed the loading parameters to vary freely and did not
specifically test for measurement invariance. Model fit was good in all samples.
The fit statistics in each sample were as follows: grade-11 males: CFI: .983,
RMSEA: .036 (.036 - .037), grade-12 males: CFI: .980, RMSEA: .040 (.040 - .041),
grade-11 females: CFI: .984, RMSEA: .034 (.034 - .035), grade-12 females: CFI:
110
.982, RMSEA: .037 (.036 - .038). Factor scores for g and the specific abilities were
saved for further analyses.
Table 5.2 Factor loadings for grade 12 males/females in the confirmatory intelligence model.
Test Name Factor
g Spatial English Speed Math
Memory for sentences .27/.35
Memory for words .49/.55 .13/.09
Disguised words .62/.64 .22/.20 .30/.31 -.16/-.15
English usage .62/.64 .41/.39
Effective expression .58/.57 .31/.24
Word functions in sent. .73/.77 .13/.06
Reading comprehension .84/.87 .11/– -.18/-.18
Creativity .71/.68 .17/.17 -.22/-.14
Mechanical reasoning .62/.62 .53/.44
Visualization in 2D .40/.41 .46/.42 .23/.18
Visualization in 3D .53/.55 .58/.54
Abstract reasoning .67/.68 .31/.28
Math 1 .79/.77 .17/.19
Math 2 .83/.73 .34/.39
Arithmetic comp. .56/.55 .34/.31 .32/.31
Table reading .22/.24 .70/.70
Clerical checking .08/.09 .73/.71
Object inspection .13/.21 .29/.25 .60/.57
Note: All freely-estimated factor loadings are shown and significant (p < .001)
111
5.3 Results
5.3.1 Factor-analytic trait complexes
Table 5.3 displays the correlations between the abilities and interests in the
grade-12 samples. They reveal that most of the interest composites were positively
related to g, except those scales relating to non-professional or semi-skilled jobs,
which had negative correlations (Trades and Clerical interests in males, and Clerical
interest in females). The scores for residual abilities displayed more differentiated,
and generally much lower, correlations with interest scales. Two notable correlations
were between Spatial ability and Science interest, and between English ability and
Arts interest. Although g and the residual abilities were orthogonal in the
intelligence model, their factor scores had slight non-zero correlations, some of
which were as large as those between interests and abilities.15
15
We removed the small amount of remaining g variance from the residual ability scores through
regression and re-ran the factor and latent class analyses. This did not substantially alter the
characters of the trait complexes obtained through either one.
112
Table 5.3 Correlations matrix for interest composites and cognitive ability factors (grade 12 males/females)
Trades Politics Sci. Cler. Med. Arts Teach. Milit. Sales Arch. g Spatial Eng. Speed Math I: Trades – .28 .49 .25 .23 .32 .23 .46 .37 – .03 .10 -.06 -.06 .01 I: Politics .07 – .44 .12 .31 .50 .49 .41 .57 – .17 -.06 -.02 .01 .07 I: Science .26 .30 – .01 .59 .45 .32 .42 .38 – .30 .09 -.08 -.04 .14 I: Clerical .33 .40 .21 – -.08 .05 .18 .13 .51 – -.20 .02 -.04 .11 .03 I: Medicine .02 .40 .47 .24 – .31 .30 .29 .23 – .16 .00 -.02 -.02 .08 I: Arts .14 .47 .34 .35 .40 – .49 .36 .53 – .31 .01 .06 -.04 -.04 I: Teaching .14 .53 .24 .45 .38 .54 – .27 .50 – .22 -.09 .03 .00 .13 I: Military .28 .36 .42 .23 .29 .29 .26 – .45 – .10 .01 -.03 -.02 .01 I: Sales .19 .63 .29 .67 .41 .56 .56 .40 – – .10 -.04 -.02 .02 .03 I: Architecture .33 .34 .49 .34 .31 .63 .31 .35 .45 – – – – – – g -.26 .14 .25 -.08 .17 .14 .11 .09 .08 .10 – .06 .07 -.01 .16 Spatial .12 -.14 .14 -.13 -.07 -.04 -.16 .01 -.18 .13 .08 – -.42 .12 -.03 English -.12 .05 -.11 .00 .04 .13 .09 -.01 .09 -.04 .13 -.34 – -.07 -.12 Speed -.04 .07 .00 .08 .05 .00 .03 .02 .06 .01 .02 .04 -.06 – -.08 Math -.05 .06 .10 .11 .03 -.07 .07 -.03 .04 -.02 .05 -.25 -.21 -.07 –
Note: Females are above the diagonal, males are below
113
We examined exploratory factor solutions for the interests and cognitive
abilities with additional factors one at a time to assess their fit and interpretability,
beginning from the one-factor solution. Model fit improved markedly from one to
three factors in all samples. Fit also improved from three to four factors, but the
four-factor solutions contained significant problems in both males and females. The
fourth factor in the grade-12 males was a near-singlet factor with Arts interest
loading above 1, and the four-factor solution did not converge in grade-12 females.
Thus, the three-factor EFA solution was used as a basis for constructing a CFA
model for the trait complexes, in conjunction with modification indices. Tables 5.4
and 5.5 display the standardized factor loadings for the CFA model in males and
females, respectively. As the aim of the trait complex model was to capture
covariance between interests and cognitive abilities, correlated residuals were
allowed if they were within the same domain (i.e. within interests or within cognitive
abilities). In the male samples there were positive correlated residuals between
Medicine and Science interests, Sales and Clerical interests, Architecture and Arts
interests, and Verbal and English residual abilities. There were negative correlated
residuals between English and Spatial ability. In the female samples there were only
two correlated residuals, positive between Military and Trades interests, and a
negative residual between English and Spatial ability. The fit statistics of the
models were only very marginally acceptable. They were as follows in each sample:
grade-11 males: CFI: .919, RMSEA: .075 (.073 - .077), grade-12 males: CFI: .914,
RMSEA: .075 (.074 - .076), grade-11 females: CFI: .907, RMSEA: .073 (.071 -
.075), grade-12 females: CFI: .908, RMSEA: .073 (.071 - .075).16
We labelled two of the factors People and Things in males and females, while
the last factor was labelled Trades in males and Clerical in females. The labels
‘People’ and ‘Things’ were inspired by Prediger (1982), who provided evidence of
two underlying dimensions in the RIASEC hexagon, one of which was termed
‘People/Things’, contrasting Social interests (People) with Realistic interests
(Things).
16
MacCallum, Browne and Sugawara (1996) characterized .080 RMSEA as “mediocre”.
114
Table 5.4 CFA solution of interests and abilities (grade 12 males)
Variable Factor
Trades People Things I: Trades .73
I: Politics .76
I: Science .70
I: Clerical .42 .55 -.17
I: Medicine .49 .10
I: Arts .51 .22
I: Teaching .71
I: Military .45
I: Sales .13 .78
I: Architecture .15 .62
g -.72 .66
Spatial -.47 .45
English -.26 .19
Speed .06
Math -.11 .07
I = interest scale score. All freely-estimated factor loadings are
shown and significant (p < .001).
In comparing the three trait complexes to those of PPIK theory, the factor
labelled ‘People’ resembled the Intellectual/Cultural trait complex and the ‘Things’
factor resembled the Science/Math trait complex. The People factor had loadings
from Artistic interests and the residual English (Verbal) ability as in Ackerman
(1996). However, in males particularly there was a greater emphasis on interest
scales reflecting Social or Enterprising interests in RIASEC terms (loadings for
Politics, Teaching and Sales) than would be predicted for the Intellectual/Cultural
trait complex. Therefore, this factor appeared to reflect a broader orientation towards
occupations involving interaction with other people (and was thus labelled the People
factor). The Things factor had loadings from Science interest and Spatial ability that
were consistent with the Science/Math trait complex (Ackerman, 1996). In the
males, Spatial ability had the highest positive loading of any residual ability.
However, this factor also appeared to be somewhat broader in scope than the
Science/Math trait complex, particularly in females where Trades interest also loaded
115
moderately on this factor, consistent with an orientation toward jobs involving
manipulation of the physical world (aligning with the Things pole of Prediger’s
dimension). The moderate loading of Medicine interest on the Things factor in
females may seem to contradict this interpretation. However, Science interest had a
loading of near-unity on the factor, and the highest correlation of Science interest
was with Medicine interest (r = .59); thus, the loading for Medicine may have been at
least partly an indirect effect of Science interest. This factor also received lower
loadings from Math ability than in PPIK theory (Ackerman, 1996).
Table 5.5 CFA solution of interests and abilities (grade 12 females)
Variable Factor
Clerical People Things I: Trades .27 .50
I: Politics .73
I: Science .99
I: Clerical .83
I: Medicine .61
I: Arts .72
I: Teaching .64
I: Military .41 .19
I: Sales .46 .67
I: Architecture
g -.31 .37
Spatial -.18 .20
English .10 -.14
Speed .09
Math .14
I = interest scale score. All freely-estimated factor loadings are
shown and significant (p < .001).
The final factor did not resemble the trait complexes proposed by Ackerman
and colleagues, but instead seemed to relate to occupational prestige or level of
general intelligence. The loadings of g on these factors were moderately negative in
females, and strongly negative in males. The male factor also obtained a negative
loading from English ability. The factor was named ‘Trades’ in males and ‘Clerical’
in females due to these being the highest–loading interest scales. Trades and Clerical
were occupational interest categories requiring less skilled work and which had
116
lower prestige than the other categories. Although not shown, the structure of the
trait complexes was highly similar in the grade-11 samples. The only notable
difference was that in the grade-11 samples the factor loadings were marginally
lower and the latent classes were slightly less distinct. However, all factor loadings
were close in magnitude to those in the grade-12 samples. In males, the loading with
the largest difference from grade 12 was for g on the Trades factor, which was .14
more positive (loading = -.58 in grade 11). In females, the largest difference was for
the loading of Sales on the Clerical factor, which was .05 less positive (loading = .41
in grade 11).
5.3.2 Latent class trait complexes
Latent class analysis was applied to the same interest and cognitive ability
scores. The number of classes was decided by examining the changes (decreases) in
Akaike information criterion (AIC; Akaike, 1983) and Bayesian information
criterion (BIC; Adrian E Raftery, 1995) of the models as additional classes were
added, as well as by considering the classification quality metric of entropy
(Ramaswamy, DeSarbo, Reibstein, & Robinson, 1993).
Examination of the AIC and BIC values showed that they exhibited an
“elbow”, or levelling off, at five classes in the male samples, and at six classes in the
female samples. At this number of classes entropy values were also remained
acceptable (close to .80), and the probabilities for most likely class membership were
.79 or greater for every class. Thus, We decided to retain five classes in males and
six in females. Entropy values were as follows in each sample: grade-11 males
(0.762), grade-12 males (0.761), grade-11 females (0.773), grade-12 females (0.767).
Tables 5.6 and 5.7 display the mean standardized values for the interests and
cognitive abilities in each latent class in grade-12 males and females, respectively.
The latent class means were highly similar in the grade-11 samples, and furnished
the same interpretations of the classes; thus they are not shown. The classes varied
widely in their mean scores, but the most notable pattern was that two classes
contained either people with low occupational interests on all scales (Class 1), or
people with high interest on all scales (Class 5). Moreover, mean g scores were
below average in the low-interest class, whereas g level was above the mean in the
117
high-interest class (females) or at mean level (males). This finding was a recurrence
of the positive correlation of most of the interest scales with g, and the factor-analytic
trait complex that related g to occupational level or prestige. The remaining classes
resembled the factor-analytic trait complexes of People (Class 3 in males, Class 6 in
females) and Things (Class 4 in males and Class 3 in females), but this distinction
was generally less clear than in the factor-analytic trait complexes. The male results
also provided a clearer separation of these two classes than the female samples. As
in the factor-analytic trait complexes, the Spatial and English residual abilities
sometimes showed an opposing pattern; for example, in grade-12 males Spatial
ability was above the mean for the Things-oriented class, but below the mean for the
People-oriented class. The Science/Math and Intellectual/Cultural trait complexes
could be identified with the Things and People-oriented classes, but as in the factor
analysis their interest associations were broader than would be anticipated based on
PPIK theory (Ackerman & Heggestad, 1997).
Table 5.6 Latent class means from LCA (grade 12 males)
Variable Latent class
1 2 3 4 5 I: Trades -.78 -.42 .17 .43
I: Politics -1.73 -1.03 .41 .31 1.46
I: Science -1.62 -.18 -.91 .61 .76
I: Clerical -1.60 -.72 .47 .17 1.15
I: Medicine -1.32 -.49 -.21 .37 .89
I: Arts -1.87 -.80 -.23 .42 1.54
I: Teaching -1.55 -.91 .42 .18 1.46
I: Military -1.46 -.24 -.34 .36 .74
I: Sales -2.84 -1.28 .73 .40 1.97
I: Architecture -2.02 -.39 -.73 .61 1.11
g -.45 .24
Spatial .31 -.90 .29 -.26
English -.17 .46
Speed
Math -.20
Note: Class means between -.14 and .14 not shown. Composition of sample: class 1 =
8.4%, class 2 = 26.0%, class 3 = 15.1%, class 4 = 34.4%, class 5 = 16.0%.
118
Table 5.7 Latent class means from LCA (grade 12 females)
Variable Latent class
1 2 3 4 5 6
I: Trades -1.04 -.35 2.33 -.46 2.32
I: Politics -1.38 -.38 -.50 2.02 1.69
I: Science -1.27 -.60 1.06 .72 1.89 .80
I: Clerical -.46 .60 .27 -1.12 .90
I: Medicine -.92 -.47 .32 .86 .96 .58
I: Arts -1.33 -.16 .38 1.25 1.02
I: Teaching -1.07 1.25 1.02
I: Military -.95 -.22 .77 -.20 1.47 .55
I: Sales -1.74 .43 .38 -.99 2.05 1.19
g -.19 .33 .92 .42 .82
Spatial .40 .19
English -.20 -.20
Speed
Math .26 .15 .20
Note: Class means between -.14 and .14 not shown. Composition of sample: class 1 =
21.4%, class 2 = 25.8%, class 3 = 12.0%, class 4 = 13.3%, class 5 = 7.6%, class 6 = 19.8%.
5.3.3 Prediction of occupational type
In the last stage of the analysis, the cognitive ability and interest scores were
used to predict the occupational type of participants eleven years after high school.
Tables 5.8 and 5.9 display the grade-12 results for the odds ratios for the logistic
regressions of the individual scores and factor-analytic trait complexes (entered in
two separate analyses). All predictors were standardized, hence the odds ratios
represent the increase/decrease in the odds of attaining the particular occupation type
given a one standard deviation increase in the variable. Due to the large sample
sizes, confidence intervals for odds ratios were very small and are not shown.17
To
examine the predictive validity of the individual scores and trait complexes, the sizes
of the odds ratios and pseudo R2 values were compared. The results for the grade-11
samples were similar and not shown.
17
For example, in the prediction of science jobs in grade-12 males, the mean 95% confidence interval
occurred from .022 below the odds ratio estimate (SD = .017) to.023 above the estimate (SD= .017).
119
The odds ratios for the individual scales were consistent with the previous
research on the predictive validity of cognitive abilities and interests (Schmidt &
Hunter, 2004; Wai et al., 2009). According to the odds ratios, the strongest cognitive
predictor for most categories was general intelligence, but the residual abilities also
made some notable contributions. For example, the English residual ability had
strong associations with the Fine Arts categories in both males and females. The
residual abilities also showed discriminative predictive validity, where higher Spatial
ability, for example, contributed positively to scientific and technical jobs, but led to
lower probabilities of attaining social-oriented jobs (such as Teaching, and
Humanities in males). The interest composites also showed good discrimination
among categories; for example, interest in Teaching was strongly predictive of
attaining a job in that category, but negatively related to attaining jobs in several
other categories.
The odds ratios for the trait complexes were generally consistent with their
effectiveness in capturing the shared variance between the interests and cognitive
abilities. The Things factor had large effects on the probabilities of entering in
scientific and technical jobs, whereas the People factor affected jobs in social-
oriented categories. Greater scores on the Trades and Clerical factors decreased the
odds of being in the professional job categories, and increased the odds of being in
the non-professional and semi-skilled job categories (such as Construction jobs in
males, and Clerical jobs in females).
120
Table 5.8 Odds ratios of abilities and interests predicting job categories (grade 12 males).
Predictor Job Category
Science Med. Business Teaching Humanities Fine Arts Technical Sales Mech. Clerical Construc. Labour
g 3.71 3.78 1.25 1.49 2.66 .81 .72 .61 .48
Spatial 1.32 .81 .69 1.34 .83 1.23
English .76 1.20 1.67
Speed 1.45 .75 1.46
Math 1.39 1.63 .70 .82 .83
I: Trades .65 .66 .64 .86 .82 1.55 1.88 1.51
I: Politics 1.21 1.30 1.17 .82
I: Science 2.41 .59 .72 .84 1.49
I: Clerical 1.19 .63 1.36 1.42 1.79
I: Medicine 2.92 .85 .68 .82
I: Arts .76 .83 1.17 1.80 1.22 .68 .82 1.36 1.22
I: Teaching .84 .79 2.66 1.19 .52 .82 .82 .80 .76 .72
I: Military 1.19 .79
I: Sales .74 1.28 1.45 .83 .61 1.38 .82
I: Arch. 1.21 .84 .80
F1: Trades .27 .19 .58 .80 .20 .47 .70 2.52 1.66 3.46 3.87
F2: People .44 1.37 2.33 2.28 .59 .43 1.20 .66
F3: Things 5.72 2.97 1.40 .67 1.48 1.60 1.89 .54 .65 .29
R2: full .52 .47 .11 .25 .42 .21 .10 .06 .22 .15 .26 .25
R2: factors .37 .39 .11 .15 .46 .11 .06 .00 .29 .06 .29 .32
Note: Odds ratios between .86 and 1.14 not shown. R2: full was for individual scales as predictors; R
2: factors was for the trait-complex factors. Med.
= Medicine, Mech. = Mechanical.
121
Table 5.9
Odds ratios of abilities and interests predicting job categories (grade 12 females).
Predictor Job Category
Science Med. Business Teaching Humanities Fine Arts Technical Sales Mech. Clerical Construc. Labour
g 3.85 1.94 1.74 2.97 .64 1.59 .70 .58 .41 .50
Spatial 1.17 .82 1.18 1.16
English 1.23 1.17 1.49 1.35 1.26 1.17 1.35
Speed 1.27 1.35 .78 1.19
Math 2.13 1.25 1.28 1.46 .67 .75 .85 .24 .77
I: Trades .61 .85 1.18 .27 1.32 .39 1.23
I: Politics 1.30 1.28 1.24 1.70 .85 .75 .75 .43
I: Science 3.92 1.31 1.20 .78 1.73 1.58
I: Clerical .66 .47 .56 2.17 .75 1.35 .47
I: Medicine .58 3.22 .64 1.32 .79 .63
I: Arts .55 .84 1.13 1.50 3.31 1.26 1.19 .74
I: Teaching .85 .79 .64 1.89 .82 .73
I: Military 1.13 .76 1.45 1.19 .69 1.52
I: Sales .81 .85 1.64 1.72 .75 1.35 .90 1.20 2.07
F1: Clerical .43 .46 1.52 .40 .28 .28 1.57 1.42 1.61 .56 1.33
F2: People .65 3.25 6.62 7.21 .60 .39 .70 .40
F3: Things 5.71 2.84 1.18 .46 .35 2.16 .65 1.77 .53 1.61
R2: full .55 .39 .13 .29 .45 .52 .21 .13 .19 .08 .63 .18
R2: factors .45 .30 .05 .29 .42 .42 .11 .07 .13 .06 .30 .11
Note: Odds ratios between .86 and 1.14 not shown. R2: full was for individual scales as predictors; R
2: factors was for the trait-complex factors. Med.
= Medicine, Mech. = Mechanical
122
The variances explained by the individual scale scores and the trait
complexes were equal or nearly equal for some categories (such as Humanities in
males and females, and Teaching in females), but were generally lower for the trait
complexes. For the grade-12 males, the mean pseudo-R2
for the trait complexes was
21.8% (SD = .15), compared with 25.2% (SD = .15) for the full scores. In females it
was 22.6% (SD = .15) compared with 31.2% (SD = .19) for the scale scores.
Therefore, our hypothesis that the trait complexes would show predictive validity
equal to the individual scores was not supported. The same conclusion was drawn
for the grade-11 results (see Supplemental Tables C1 and C2).
Tables 5.10 and 5.11 display the odds ratios from logistic regressions of the
jobs categories onto latent class memberships in grade-12 males and females. The
reference class was chosen as the largest group (class 4 in males and class 2 in
females). The odds ratios in the male data were generally smaller than for the
individual scores or factor-analytic trait complexes. However, in grade-12 females,
several large odds ratios were observed for the probabilities of attaining Science,
Medicine and Fine Arts jobs. This was likely due to the small frequencies of jobs in
these categories for females, such that those who attained them were outliers in
interests and abilities. The mean pseudo-R2 for the LCA trait complexes was 6.3%
(SD = .04) for grade 12-males, and 13.7% (SD = .09) for the grade-12 females.
These values were considerably lower than the variance explained by the scale scores
(the same pattern occurred in the grade-11 samples; see Supplemental Tables C3 and
C4). The higher explained variance in females was likely partially attributable to the
use of one additional class. The hypothesis of equal predictive validity was clearly
rejected for the latent-class trait complexes.
123
Table 5.10 Odds ratios of latent classes in predicting job category (grade 12 males).
Predictor Job Category
Science Med. Business Teaching Humanities Fine Arts Technical Sales Mech. Clerical Construc.
a
Labour
Class 1 .33 .16 .37 .64 .39 1.89 .76 1.80 .31 1.68 2.96
Class 2 .69 .41 .47 .74 .80 1.33 1.26 1.54 2.05
Class 3 .10 .57 3.60 2.30 .50 .28 .45 1.17 .51 1.37
Class 5 .63 .40 2.10 1.97 1.76 .67 .68
R2 .15 .08 .05 .10 .07 .05 .06 .00 .05 .04 .04 .04
Note: Reference class is class 4. Odds ratios between .86 and 1.14 not shown. Med. = Medicine, Mech. = Mechanical
Table 5.11 Odds ratios of latent classes in predicting job category (grade 12 females).
Predictor Job Category
Science Med. Business Teaching Humanities Fine Arts Technical Sales Mech. Clerical Construc.a Labour
Class 1 3.90 1.57 .52 .78 .34 3.36 .44 3.24 .73 n/a 1.99
Class 3 9.56 2.47 2.57 1.41 1.39 1.60 .71 1.25 1.41 .66 n/a 2.27
Class 4 15.60 15.11 .35 2.40 1.58 3.38 .29 .12 2.91 .46 n/a
Class 5 11.89 6.26 .77 2.82 1.36 .32 1.34 n/a
Class 6 18.05 4.90 1.16 2.94 2.83 25.52 .39 .61 n/a .24
R2 .29 .20 .09 .08 .13 .28 .06 .14 .08 .02 n/a .14
Note: Reference class is class 2. Odds ratios between .86 and 1.14 not shown. Med. = Medicine, Mech. = Mechanical
a model did not converge.
124
5.3.4 Multinomial prediction
An additional way to predict the occupational categories in Project TALENT
was using multinomial logistic regression, where the binary outcomes of the twelve
categories were predicted simultaneously. An advantage of this method was that the
ability to classify individuals into the correct occupation category could be
determined, comparing across the three different sets of predictors (the individual
scores, the factor-analytic trait complexes, and the latent-class trait complexes).
Austin and Hanisch (1990) similarly examined the classification accuracies of their
five discriminant functions in the grade-10 sample of Project TALENT; therefore our
results can also be compared to theirs.
Tables 5.12 and 5.13 display the correct classification percentages from the
multinomial regressions for the grade-12 males and females (the analysis was not
performed for the grade-11 samples). The percentages are provided for the three
different sets of predictors. The correct classification percentage was greater than
chance when it exceeded the sample percentage (the percentage in the population
who were in that occupation category); hence these values are given for comparison.
The sample percentages differ slightly from those in Table 5.1 because the “Vague
and Undesigned” occupation category was excluded (since it would not be expected
to be able to classify individuals in that category).
The mean classification accuracies are given in the bottom rows. The mean
accuracies indicated that the individual scores had the highest classification accuracy,
followed by the factors and then the latent classes. The classification accuracy
exceeded chance with the individual scores for nine of the twelve categories in males
and six of twelve in females. This accuracy was reduced when the factors or classes
were used as predictors. This finding replicated the results from the logistic
regression that the factors and classes had lower predictive power than the individual
scores for cognitive abilities and interests.
125
Table 5.12
Original sample composition and correct classification percentages for multinomial
regression (grade 12 males)
Classification success (% correct)
Job category Sample
percentage
Individual Scores Factors Classes
Science 6.2 42.0 17.8 0
Medicine 2.8 16.4 0 0
Business 21.3 62.0 74.2 79.0
Teaching 9.1 24.4 1.9 0
Humanities 3.9 16.5 3.0 0
Fine Arts 1.0 0 0 0
Technical 5.7 0.4 0 0
Sales 11.9 9.1 0 0
Mechanical 9.6 20.0 15.5 0
Clerical 8.5 19.8 1.4 0
Construction 3.8 3.0 0 0
Labour 16.2 61.0 61.6 49.1
Mean 8.33 22.9 14.6 10.6
Table 5.13 Original sample composition and correct classification percentages for multinomial
regression (grade 12 females)
Classification success (% correct)
Job category Sample
percentage
Individual Scores Factors Classes
Science 0.4 17.6 0 0
Medicine 7.3 35.3 3.5 0
Business 7.3 0.3 0 0
Teaching 19.2 55.8 42.4 20.2
Humanities 2.1 1.7 0 0
Fine Arts 1.0 27.8 0 0
Technical 4.5 0 0 0
Sales 4.5 0 0 0
Mechanical 1.1 0 0 0
Construction 0.04 0 0 0
Clerical 34.6 74.7 80.0 92.0
Labour 18.0 38.7 17.2 0
Mean 8.33 21.0 11.9 9.4
5.4 Discussion
Cognitive abilities and occupational interests are intertwined, and several
developmental theories have been advanced to explain these associations (Ackerman,
1996; Gottfredson, 1986; Hogan & Roberts, 2000). Only PPIK theory, however, has
provided hypotheses of the overlap between particular interests and abilities, which
are said to be captured by two trait complexes termed Math/Science and
126
Intellectual/Cultural (Ackerman, 1996; Ackerman & Heggestad, 1997). Trait
complexes have been proposed as important influences on occupational knowledge,
and thereby career choice (Ackerman, 1996; Ackerman & Beier, 2003a). Yet,
previous research had not examined whether trait complexes can predict future
occupation. In addition, Ackerman and colleagues have relied on factor analysis to
extract trait complexes, though latent class analysis is arguably more consistent with
their definition. We reasoned that if PPIK theory was correct then trait complexes
obtained from interest and cognitive ability scores in Project TALENT would fulfil
two conditions. First, the content of the Science/Math and Intellectual/Cultural trait
complexes would be replicated, and second, they would show equal or greater
predictive validity than individual scales score for predicting occupational type. The
first condition received only mixed support and the second was not supported; we
discuss each of the hypotheses in turn.
When an acceptable confirmatory factor analytic model of the trait complexes
was constructed, factors were found that resembled the Science/Math and
Intellectual/Cultural trait complexes. However, the involvement of interests in the
trait complexes was broader than proposed in PPIK theory, and the factor content
aligned more closely with the two poles of the People/Things interest dimension
(Prediger, 1982). The People factor had loadings from Sales, Politics and Teaching
interests, which correspond to Enterprising and Social interests in the RIASEC
framework. These interests were not hypothesized to be part of the
Intellectual/Cultural trait complex (Ackerman & Heggestad, 1997), but are
characteristic of the People interest pole (Prediger, 1982). The Things factor
displayed a closer correspondence with the Science/Math trait complex, except that
the loading of residual Math ability was minimal. Johnson and Bouchard (2009) also
found that a broad section of people-oriented interest groups had higher verbal
abilities, while groups that were things-oriented displayed higher spatial (image
rotation) abilities. This also contradicted the more narrow focus of PPIK theory.
However, in this study, the most notable departure from PPIK theory was the
presence of a third factor.
127
The factor captured the negative associations of general intelligence with the
Trades and Clerical interests. Tracey and Rounds (1996) found a third dimension of
the RIASEC interests that is relevant to this finding. The first two dimensions were
defined by Prediger (1982), who labelled them People/Things and Data/Ideas
(Data/Ideas was oriented between Conventional and Enterprising interest (Data) and
Investigative and Artistic interests (Ideas). Tracey and Rounds (1996) performed a
principal component analysis of interest ratings for 229 occupation titles, and found a
third component that was related to occupational prestige or socioeconomic status.
Our third factor was consistent with such a prestige dimension, and specifically
included g, which Tracey and Rounds did not measure. The involvement of g in the
trait complexes was extensive. Considered together, the trait complexes explained
by far the most variance in the g factor. For example, in the grade-12 males, the trait
complexes explained 43.7% of the variance in g, but only 18.9% for the next-highest
ability (Spatial ability). In fact, there was more variance explained in g than in the
four specific abilities combined, which had a total of 27.6% of their variance
explained by the trait complexes. In the grade-12 females, this distinction was even
stronger, with 21.9% of the g variance explained, and 7.2% for the specific abilities
combined. These ratios were similar in the grade-11 samples. In the latent class
analysis, it was observed that classes which reflected greater interests in higher-status
occupations (such as Science or Politics) also had above-average g levels. General
intelligence was notably involved with two of the five LCA trait complexes in males,
and five of six in females, while the specific abilities generally played lesser roles.
In the PT scales, there was a moderate general interest factor, and it was
associated with g, particularly according to the LCA results. Some researchers have
cautioned that the general interest factor is likely a “response set” owing to
acquiescence bias or other methodological factors, and have advised to control for it
(e.g. Prediger, 1982). Some occupational interest scales intrinsically control for
differences in average level of response by ipsatisation. However, other researchers
have suggested that the general interest level could have some substantive
psychological meaning, noting that it correlates positively with Extraversion and
Emotional Stability (see Rounds & Tracey, 1993, and the references therein).
128
The Science/Math and Intellectual/Cultural trait complexes of PPIK theory
emphasize the importance of specific abilities in interest-ability associations, but we
found that g played a more important role in the trait complexes than did specific
abilities. These findings replicated those of Johnson and Bouchard (2009) that there
were substantial differences in mean g level across latent-class interest groups, in line
with their average occupational status. The results lend support to Gottfredson’s
theory of circumscription and compromise, which specifies that g plays a central role
in determining the occupations in which individuals become interested, according to
the levels of education and training required (Gottfredson, 1986, 2005).
PPIK theory specifies the involvement of intelligence-as-process in the
Science/Math trait complex (where intelligence-as-process could be considered
similar to g), but does not propose direct involvement of consideration of social
status or training requirements in the emergence of the trait complexes. One possible
reason for this is that Ackerman and colleagues have used RIASEC measures of
interests, which are limited in their representation of low-prestige occupations (Deng
et al., 2007; Tracey & Rounds, 1996). In contrast, the Trades and Clerical interest
scales derived from PT items were primarily formed from low-prestige occupation
titles. Moreover, as noted in the introduction, Ackerman and colleagues did not
separate g variance from specific-ability variance in measuring cognitive abilities,
and thus were not able to evaluate the roles of g and specific abilities separately in
their trait complexes. In addition, they have often used samples of college students,
which suffer from range restriction of cognitive ability, as well as occupational
interests (Ackerman, 2000; Kanfer, Wolf, Kantrowitz & Ackerman, 2010).
The People and Things factors were consistent with the People/Things
interest dimension. However, the factors were positively correlated, which suggests
that they did not act as poles of one dimension in the current study. This may be
attributable to differences in the interest scales used. Studies that assess the RIASEC
types typically find that the correlations among types vary from moderately positive
to moderately negative (De Fruyt & Mervielde, 1999). In the PT data, all the interest
composites correlated positively. This shared variance between interest scales made
it unlikely to obtain factor-analytic trait complexes would that were negatively
129
correlated with each other, given that the interest scales made up the majority of the
variables entered in the analysis.
One source of the more positive correlations among the PT scales compared
with the RIASEC measures could have been that the RIASEC scales were designed
to emphasize the separation between the types by selecting occupation titles that are
unambiguous representatives, whereas the PT occupation titles were not pre-selected
to fit separate categories. Thus, a substantial portion of each item response was
made up of the student’s general level of occupation interest. Another
methodological factor that could have contributed to this common variance was
acquiescence due to testing fatigue because the participants were young and were
required to complete many scales during the course of the study (Flanagan et al.,
1962). Youth could have contributed substantively to the general interest factor as
well, given that the students may not have been aware of the challenges in different
occupations and thus responded more positively to a wide variety of titles than older
and more knowledgeable respondents would have done. Thus, the data in the current
study were probably not ideal to assess whether People/Things consists of one bi-
polar dimension or two dimensions, although the results did suggest two separate
dimensions. Overall, the People and Things factors that we observed were generally
consistent with Intellectual/Cultural and Science/Math trait complexes, but not
identical to them. While the cognitive abilities generally showed the expected
relations with the two factors, the interest loadings appeared to capture divisions that
were more consistent with Prediger’s People/Things distinction than the distinction
between Cultural and Scientific interests emphasized in PPIK theory.
The structures of the trait complexes using both methods were very consistent
across grades. The structure was slightly less clear structure in grade 11, but this
could have simply resulted from the interests and cognitive abilities being less
developed and differentiated in the younger sample. The trait complexes were less
consistent between genders than grades, but in both groups three factors were
obtained that were recognizable as People, Things and prestige factors. In the latent
class analysis, the numbers of classes obtained for males and females differed by
one, but both sets of classes captured groups that were organized according to
130
occupational level and the People-Things distinction. The prestige factor for females
primarily related g negatively to Clerical interests, while in males it related g
negatively to Trades interests. This difference is consistent with the finding that
males are on average more interested in Realistic-type occupations which Trades fall
under, while females are more interested in Social-type interests, which are relevant
to the Clerical factor (Deng et al., 2007). However, social roles for men and women
were also likely involved in this difference in which lower-prestige occupations they
preferred. In the 1970’s it was very uncommon for women to enter Realistic-type
occupations, and less likely for men to enter Social-type occupations that were
Clerical in nature, compared to other non-Clerical jobs. The limitation of the time
period in regard to sex differences is addressed further below. In summary, the trait
complexes obtained were consistent across samples yet differed from those predicted
in PPIK theory. The present study went beyond previous research in studying the
associations of interests to cognitive abilities, to the prediction of attained
occupation. The individual scores for cognitive abilities and interests had substantial
power in predicting occupational category eleven years after high school, consistent
with previous studies (J. T. Austin & Hanisch, 1990; Humphreys et al., 1993).
However, this was the first study using PT data to find that specific abilities,
independent of g, also predicted some occupations. Most notably, residual Spatial
and English abilities displayed the clearest patterns of prediction, with Spatial ability
predicting scientific and technical jobs and English ability predicting jobs in the
broad Humanities area. Residual Math and Speed abilities were less consistent
predictors, but Math ability, for example, was predictive of future Science jobs in
both males and females.
Trait complexes derived by factor analysis were strong predictors of
occupational type, but, excepting a few occupational categories, they explained less
variance than individual scores for cognitive abilities and interests. This was the
case using both logistic and multinomial regression. The latent-class trait complexes
performed notably worse than the factor-analytic trait complexes in predicting
occupational type.
131
The weaker predictive ability of latent-class trait complexes suggests that trait
complexes, if they exist, should not be conceptualized as groups of individuals who
share similar levels of interest and cognitive ability variables. Rather, there was
greater support for the idea that trait complexes could be conceptualized as capturing
parts of the shared variance between cognitive abilities and interests, where these
variables and the trait complexes are continuous. This shared variance may be the
result of the reciprocal influences of cognitive abilities and interests upon each other
through development, as theorized by a number of researchers (Ackerman, 1996;
Armstrong et al., 2008; Hogan & Roberts, 2000), although direct evidence for this is
still lacking.
One possible limitation of the current study was that the trait complexes that
primarily involve personality-intelligence associations were excluded (the Social and
Clerical/Conventional trait complexes). Nonetheless, Ackerman and Beier (2003)
put forth the general claim that trait complexes are more informative about career
choices than individual scales, which should apply to all their trait complexes. In
addition, previous research on trait complexes has found the Science/Math and
Intellectual/Cultural trait complexes to be the two most important predictors of
specialized knowledge (Ackerman & Rolfus, 1999) and the university course that
students select (Ackerman, 2000). Nonetheless, future research could be done to
address the predictive validity of the Social and Clerical/Conventional trait
complexes for attained occupation.
A second limitation was that the selection of the number of trait complexes
through EFA and LCA was subjective, and guided in large part by their
interpretability. The confirmatory models also displayed marginal fit, although the
use of CFA is an advance over previous studies of trait complexes that have only
used exploratory methods. The difficulty in constructing factor models for the trait
complexes may have been exacerbated by problems with the interest and cognitive
ability measures. As discussed above, the interest scales displayed a positive
manifold, which is inconsistent with research on the RIASEC that has found a
circumplex structure for interests (Armstrong et al., 2008; Holland, 1997). The
amount of variance for which the residual abilities accounted was small, which was
132
at least partly attributable to the modest number of tests for each. This may have
contributed to the weaker involvement of the residual abilities in the trait complexes
compared to g, as well as their lesser predictive validity. However, previous studies
have also found that g has the greatest importance (Gottfredson, 1986; Johnson &
Bouchard, 2009).
Another limitation of research on future occupation is that the power of
prediction is dependent on the job market. If the skills and interests in the population
do not match the requirements for the available jobs, at least some individuals will be
mismatched. In addition, there are social and economic pressures that may act to
lead individuals away from their ideal occupations. The 1970s was a period of
increasing educational opportunity in the United States, but the occupational
opportunities were not as dominated by educational qualifications as in the present
day. For women, strong social expectations about gender roles in the division of
labour were present: nearly half the women at follow-up were housewives, and the
most prevalent paid occupation was in the Clerical and Office Work category.
Gender expectations prevented many women from selecting jobs for which their
cognitive abilities and interests were suited. However, the total predictive validities
of abilities and interests were not substantially lower for women compared with men,
even for male-dominated occupation categories. This was possibly because men
with a wider range of abilities and interests would have obtained these jobs, thus
diluting the predictive validity of the baseline variables. One notable divergence was
that the explained variance was greater for Fine Arts occupations for females than
males, which suggests that gender roles may have also restricted the occupational
opportunities of men. In comparison with the present day, manufacturing and trades
jobs were more prevalent, which provided a niche for more low-g workers; the
predictive validity of g and trades’ interests may be lower in modern samples relative
to PT.
The mean classification accuracy across the twelve categories in Austin and
Hanisch (1990) was 30.5%. Here it was lower, even with the use of full scores
(22.9% in males, 21.0% in females). There were several factors that likely
contributed to this. First, Austin and Hanisch did not separate their sample by
133
gender, and instead entered it as a variable in the discriminant function analysis. As
the genders differed strongly in their frequencies across occupation category (see
Table 5.1), this would have increased their predictive power compared to ours.
Socioeconomic status was also used as a predictor in that study but not in the current
one.
In addition, a limitation of multinomial regression is that unequal proportions
in the outcome variable decreases the prediction accuracy because individuals are
more likely to be classified into the more common prior categories. For example, the
two most common categories in males, Business and Labour, were overrepresented
in the classifications. The grade-10 sample of Project TALENT was more evenly
distributed amongst the twelve categories than the grade-11 or grade-12 samples,
which likely contributed to the greater classification accuracies found by Austin and
Hanisch (1990).
In summary, our first finding was that the Science/Math and
Intellectual/Cultural complexes in PPIK theory could not be closely replicated
because the trait complexes we found were broader in content and gave much more
weight to g. Within the factor-analytic trait complexes, g had the most explained
variance, and one factor related low g to interest in Clerical and Trades occupations
(identified with the prestige dimension of Tracey and Rounds [1996]). These
findings were consistent with the theory of Gottfredsson (1986, 2005) that g acts as
an important filter in occupations according to status level. Neither type of trait
complexes were equal predictors of attained occupation when compared to individual
traits, which calls their theoretical status into question. The greater predictive
validity for factor-analytic trait complexes than latent-class trait complexes suggests
that the Science/Math and Intellectual/Cultural trait complexes, if their definition is
expanded, may be useful summaries of the overlap between cognitive abilities and
interests, but they do not appear to represent discrete groups in the population with
combinations of different trait levels.
134
Chapter 6: Conclusion
The studies in this thesis examined the links among cognitive ability, personality
and occupational interests in Project TALENT. The research built towards testing of
a key component of the integrative framework of Ackerman and colleagues
(Ackerman, 1996; Ackerman & Beier, 2003a; Ackerman & Heggestad, 1997), the
concept of trait complexes made up of cognitive abilities and interests. In this final
chapter, the main results of the studies are summarized, and their implications are
discussed within the context of research that strives to integrate the three domains of
individual differences. Some of the limitations of PT data to address this topic are
discussed, and suggestions for future research are provided.
5.1. Cognitive ability
The study presented in chapter 3 was designed to investigate the
psychometric structure of the cognitive ability tests in PT. Three of the most well-
regarded models were compared: the VPR model, the CHC and the Extended Gf-Gc
models. The VPR model was found to have the best fit to the test data in all samples.
The results provided replication of three previous model comparison studies where
the VPR model outperformed the CHC and Gf-Gc models (Johnson & Bouchard,
2005a, 2005b; Johnson, Te Nijenhuis, et al., 2007). This comparative research has
suggested that the VPR model is the most accurate representation of human cognitive
abilities, thus it follows that this model should be the best suited for understanding
the overlap among cognitive abilities, personality and interests. Nonetheless, the
main purpose of this study in the thesis was to develop a model appropriate for use in
summarizing the cognitive ability measures in PT for further research. The topic of
how the VPR model can contribute to integrative research is discussed in section 5.3
below.
5.2. Personality-intelligence associations
The study presented in chapter 4 investigated linear and nonlinear
associations between general intelligence and personality. The linear associations
that were observed were in line with previous research: g was positively associated
135
with PT scales reflecting Openness to Experience and negatively related to
Neuroticism scales (Ackerman & Heggestad, 1997; Zeidner and Shani-Zinovich,
2011). There was mixed support for the hypothesis of a negative association
between g and Conscientiousness: g was negatively associated with Tidiness, but not
Maturity (Chamorro-Premuzic & Furnham, 2006). Finally, the results supported
Ackerman and Wolf’s (2005) hypothesis that the social potency aspect of
Extraversion is positively associated with g, while social closeness is negatively
associated.
In contrast to most previous studies of nonlinear associations, several
significant quadratic effects of g on personality traits were also found. These
quadratic associations were predicted primarily on the basis of previous research
with gifted (high g) samples (Sak, 2004; Zeidner & Shani-Zinovich, 2011). Three of
the most consistent nonlinear associations were between greater g and greater social
potency (Leadership), lower social closeness (Sociability), and lower scores on the
Tidiness facet of Conscientiousness. Another conclusion drawn from the study was
that the general factor in personality self-ratings may be an important confound to
consider when studying personality-intelligence associations.
These findings can be applied to integrative research. PPIK theory and the
integrative framework of Armstrong and colleagues have only addressed linear
associations between cognitive abilities and personality (Ackerman & Heggestad,
1997; Anthoney & Armstrong, 2010; Armstrong et al., 2008). This limitation is due
to the methods used to identify associations between the three domains: factor
analysis and multidimensional scaling, which only capture linear associations among
variables. However, even within linear associations, neither theory has taken into
account the negative associations of g with Neuroticism and the Tidiness aspect of
Conscientiousness, as well as the differential link of g with the two aspects of
Extraversion. These links, although smaller than the relation between g and
Openness to Experience, could be important in understanding how g transacts with
personality to influence the development of occupational interests. For example,
individuals with higher g scores may be less interested in occupations in the
Conventional RIASEC domain, not just because of the occupations’ lower average
136
prestige, but because Conventional jobs have a high requirement for
Conscientiousness, which includes Tidiness (Armstrong et al., 2008).
Although the methods used by existing theorists have not been suited to
incorporating nonlinear associations, one possible method to do so is to examine
groups selected by extreme g scores. The results from chapter 4 suggest that a group
defined by high g would have outlying scores on the personality traits that
demonstrated nonlinear discontinuity with higher g (higher social potency, lower
social closeness and lower Conscientiousness). Understanding the set of
characteristics that are specific to groups with high g or low g levels could also be
important for an integrative theory. For example, the higher average social potency
of the intellectually gifted may incline them towards Enterprising occupations, but
their lower average social closeness may lead them away from Social occupations
that involve working personally with others. Such influences may be underestimated
if only linear associations between g and personality are considered. This conclusion
only addresses the associations between g and personality traits, but nonlinear
associations may also exist between domain-specific cognitive abilities and
personality traits. In fact, nonlinear associations may exist among all three of the
domains, but this is beyond the scope of the current analyses.
Issues surrounding the “general factor of personality” are also a concern for
integrative research. Across the eight grade and gender samples in PT, this
personality factor displayed a mean correlation of .28 with g, an association which
has also been observed in several other studies (Dunkel, 2013; Irwing, Booth,
Nyborg, & Rushton, 2012; Loehlin, 2011). In such cases, the “lower-order”
personality traits will all tend to be positively correlated with g, though the
associations may be non-significant or negative when the GFP is controlled. Future
research is needed to determine if the association between g and the GFP involves
genuine personality variance or some artifactual source, as the standing of the GFP is
still in question (Chang et al., 2012; Hopwood, Wright, & Brent Donnellan, 2011).
In light of this issue, studies that examine the associations of personality and
cognitive abilities should control for socially-desirable responding, ideally by using
multiple raters instead of social desirability scales, as social desirability scales also
137
contain substantive variance (Paulhus, 2002). The presence of a strong GFP in the
PT personality scales, with no way to control for socially desirable responding, is a
significant limitation of this dataset for integrative research.
5.3. Trait complexes
In chapter 5, the concept of trait complexes of interests and cognitive abilities
was tested by examining their long-term predictive validity. It was found that three
trait complexes obtained by factor analysis had nearly the same predictive power for
occupation as individual scores of interests and cognitive abilities, while trait
complexes obtained by latent class analysis performed substantially more poorly than
either. As latent class analysis is more consistent with the definition of trait
complexes by Ackerman and colleagues (Ackerman & Beier, 2003a; Ackerman &
Heggestad, 1997), this definition was undermined. Instead of being viewed as
combinations of levels of traits that exist in certain groups of the population, trait
complexes could only be defended as reflecting shared variance among continuous
variables. This primary conclusion is in line with the view of Armstrong and
colleagues that the integration of cognitive abilities, interests and personality is to be
best understood by considering the associations among dimensional variables
(Anthoney & Armstrong, 2010; Armstrong et al., 2008).
The People and Things factors that were obtained were broader in their
content than was predicted from PPIK theory (the Science/Math and
Intellectual/Cultural trait complexes). This result was in line with previous findings
by Johnson and Bouchard (2009) regarding the distribution of verbal and spatial
abilities across interest groups. In both that study and in chapter 5, verbal ability was
associated with broadly people-oriented interests, while spatial ability was aligned
with things-oriented interests. Moreover, in chapter 5, I found that factors
representing these overlaps were differentially predictive of technical-scientific
versus artistic-humanities jobs. When considered in the context of the VPR model,
these findings strongly suggest that the Verbal-Image Rotation dimension is aligned
with the People/Things dimension first proposed by Prediger (1982). The
Verbal/Spatial distinction in cognitive abilities may be key to understanding how the
People/Things dimension in interests emerges. Differential success in dealing with
138
verbal and spatial tasks may influence individuals’ future interest levels in People
versus Things occupations because of their different task demands. The
People/Things dimension is defined by interest in working closely with other people
versus working with physical objects and data. However, People-oriented
occupations are more likely to require verbal communication abilities, while Things-
oriented occupations are more likely to require skill at spatial manipulation. Thus,
through perceptions of their own verbal and spatial abilities, and their knowledge of
job activities, individuals are likely to gravitate towards either People or Things-type
occupations.
In addition to the important role of the Verbal-Image Rotation dimension in
the VPR model, Johnson and Bouchard (2007) found that it is the primary dimension
along which sex differences occur in cognitive abilities. This is consistent with its
association with the People/Things dimension, as the sex difference on this
dimension is one of the largest among psychological traits (Armstrong et al., 2011;
Lubinski, 2000). There is some evidence that the sex differences in both interests
and cognitive abilities could be driven by genetic differences that are traceable to
different evolutionary investment strategies for males and females (see Johnson &
Bouchard, 2009, and the references therein). However, in Johnson and Bouchard
(2009), sex differences in occupational interests were found to be larger than those
for specific and general cognitive abilites, suggesting that socialization pressures
may cause a greater separation in interests than would be expected on the basis of
cognitive abilities alone (Johnson & Bouchard, 2009).
Along with its better fit over rival models, the central role for sex differences
in cognitive ability in the VPR model is an important element that supports its use for
integrative research. In contrast, these differences are less well-articulated in
theories for the CHC model (Horn & Blankson, 2005) and Gf-Gc models (McGrew,
2009). For example, in the latest review of CHC theory, sex differences were not
even mentioned (McGrew, 2009). In these models, the distinction between spatial
and verbal abilities is typically confounded with the distinction between fluid and
crystallized factors, which undermines the investigation of sex differences. For
example, tests of crystallized intelligence that rely on general knowledge have often
139
been found to favour males (Keith, Reynolds, Patel, & Ridley, 2008). However, tests
that require specific verbal knowledge, such as spelling and vocabulary, tend to favor
females (Johnson, Bouchard, et al., 2007). Therefore, focusing on crystallized
intelligence rather than verbal intelligence may lead researchers to overlook a salient
sex difference in cognitive ability. PPIK theory still refers to the Gf-Gc model
(Ackerman, 1996), and Armstrong and colleagues have not explicitly adopted any
intelligence theory (Armstrong et al., 2008). The results of chapter 5 suggest that
both theories could benefit from incorporation of the VPR model of cognitive
abilities.
Another crucial oversight in PPIK theory is that it does not include the
prestige dimension in occupational interests and its association with g. In PT, the
involvement of g in the Science/Math trait complex was substantial, as predicted in
PPIK theory, but g was also related to greater overall occupational interests, and
particularly to interests in occupations with higher prestige. The strong association
of g with higher-prestige interests supported the developmental theory of Gottfredson
(Gottfredson, 1986, 2005).
The three “trait complex” factors found in the study tended to support the
RIASEC-based approach of Armstrong and colleagues (Armstrong et al., 2008). As
discussed above, two of the factors appeared to align with the People/Things
dimension that underlies the RIASEC (Armstrong et al., 2008; Prediger, 1982). The
third factor was related to occupational prestige or level, a dimension which has also
been incorporated into their framework (Armstrong et al., 2008; Deng et al., 2007).
However, typical RIASEC measures such as the Vocational Preference Inventory do
not contain enough breadth of occupations to cover the whole prestige dimension
(Deng et al., 2007). In addition, Armstrong and colleagues have used aptitude
ratings instead of objective intelligence tests to assess cognitive abilities. Further
research is needed to replicate the association of this third interest dimension with g.
A limitation of the study in chapter 5 was that personality was not included,
in order to focus on first integrating cognitive abilities with interests. It is possible
that the trait complexes in PPIK theory that involve personality-interest associations
could be found in PT, and that they may display better predictive validity than the
140
trait complexes for interests and cognitive abilities. However, as observed in chapter
4, the PT personality measures have greater limitations for this research than do
those for the interests, because of the lack of item-level data and the presence of a
large common factor.
5.4. Suggestions for future research
The research presented in this thesis has provided several important
indications for future integrative research. First, the studies in this thesis support the
conclusion that the VPR model of cognitive abilities is likely the best existing model
of cognitive abilities for this integrative research. This is due to its better description
of the structure of cognitive abilities, and the strong links between the Verbal-Image
Rotation dimension and the People/Things dimension of occupational interests. In
addition, future studies should separate g from specific abilities in their intelligence
models (i.e. specify them to be uncorrelated factors), as this is the only way to
accurately assess their relative contributions in integrative models.
Second, integrative research could benefit from the examination of nonlinear
personality-intelligence associations, which were responsible for some substantive
personality differences of individuals at high and low g levels. These personality
differences, such as lower social closeness and tidiness, and higher social potency,
could be relevant to the development of occupational interests, and apparently can
only be found if nonlinear models are employed. Another issue that deserves more
attention in research on personality-intelligence associations is the “general factor of
personality”, which may be an important confound in understanding these
associations.
Finally, the concept of trait complexes as originally proposed by Ackerman
and colleagues appears to be untenable (Ackerman, 1996; Ackerman & Beier,
2003a). Based on their predictive validity, the overlap among cognitive abilities,
personality and interests is best conceived as being among continuous variables. One
of the most promising approaches is to use the dimensions that underlie the RIASEC
model to anchor cognitive abilities and personality traits (Anthoney & Armstrong,
2010; Armstrong et al., 2008). The factors of interests and cognitive abilities found
in chapter 5 bore a strong resemblance to the People/Things and prestige dimensions
141
found in the models of Armstrong and colleagues. Hence, integrative models may be
best based around the structure of the RIASEC model of occupational interests,
although further research is needed to compare this approach more directly to PPIK
theory, and to examine its capacity to predict occupational outcomes.
142
References
Ackerman, P. L. (1996). A theory of adult intellectual development: Process,
personality, interests, and knowledge. Intelligence, 22(2), 227-257. doi:
10.1016/s0160-2896(96)90016-1
Ackerman, P. L. (2000). Domain-Specific knowledge as the "dark matter" of adult
Intelligence, Gf/Gc, personality and interest correlates. The Journals of
Gerontology Series B: Psychological Sciences and Social Sciences, 55(2),
P69-P84.
Ackerman, P. L., & Beier, M. E. (2003a). Intelligence, personality, and interests in
the career choice process. Journal of Career Assessment, 11(2), 205-218.
Ackerman, P. L., & Beier, M. E. (2003b). Trait complexes, congitive investment and
domain knowledge. In R. J. Sternberg & E. L. Grigorenko (Eds.),
Perspectives on the psychology of abilities, competencies, and expertise (pp.
1-30). New York: Cambridge University Press.
Ackerman, P. L., & Heggestad, E. D. (1997). Intelligence, personality, and interests:
evidence for overlapping traits. Psychological bulletin, 121(2), 219.
Ackerman, P. L., & Rolfus, E. L. (1999). The locus of adult intelligence: Knowledge,
abilities, and non-ability traits. Psychology and Aging, 14, 314-330.
Akaike, H. (1983). Information measures and model selection. Bulletin of the
International Statistical Institute, 50(1), 277-291.
Anthoney, S. F., & Armstrong, P. I. (2010). Individuals and environments: Linking
ability and skill ratings with interests. Journal of counseling psychology,
57(1), 36.
Anusic, I., Schimmack, U., Pinkus, R. T., & Lockwood, P. (2009). The nature and
structure of correlations among Big Five ratings: The halo-alpha-beta model.
Journal of Personality and Social Psychology, 97(6), 1142.
Armstrong, P. I., Day, S. X., McVay, J. P., & Rounds, J. (2008). Holland's RIASEC
model as an integrative framework for individual differences. Journal of
counseling psychology, 55(1), 1-18.
Armstrong, P. I., Su, R., & Rounds, J. (2011). Vocational interests: The road less
traveled. The Wiley-Blackwell Handbook of Individual Differences, 608-631.
Ashton, M. C., & Lee, K. (2012). On Models of Personality Structure. European
Journal of Personality, 26(4), 433-434.
Ashton, M. C., Lee, K., Goldberg, L. R., & de Vries, R. E. (2009). Higher order
factors of personality: do they exist? Personality and Social Psychology
Review, 13(2), 79-91.
Austin, E. J., Deary, I. J., & Gibson, G. J. (1997). Relationships between ability and
personality: Three hypotheses tested. Intelligence, 25(1), 49-70. doi:
http://dx.doi.org/10.1016/S0160-2896(97)90007-6
Austin, E. J., Deary, I. J., Whiteman, M. C., Fowkes, F., Pedersen, N. L., Rabbitt, P.,
. . . McInnes, L. (2002). Relationships between ability and personality: Does
143
intelligence contribute positively to personal and social adjustment?
Personality and Individual Differences, 32(8), 1391-1411.
Austin, J. T., & Hanisch, K. A. (1990). Occupational attainment as a function of
abilities and interests: A longitudinal analysis using Project TALENT data.
Journal of Applied Psychology, 75(1), 77-86.
Bäckström, M., Björklund, F., & Larsson, M. R. (2009). Five-factor inventories have
a major general factor related to social desirability which can be reduced by
framing items neutrally. Journal of Research in Personality, 43(3), 335-344.
Barrick, M. R., & Mount, M. K. (1991). The big five personality dimensions and job
performance: a meta-analysis. Personnel Psychology, 44(1), 1-26. doi:
10.1111/j.1744-6570.1991.tb00688.x
Barrick, M. R., Mount, M. K., & Gupta, R. (2003). Meta‐analysis of the relationship
between the five‐factor model of personality and Holland's occupational
types. Personnel Psychology, 56(1), 45-74.
Bartholomew, D. (1987). Latent variable models and factor analysis. London,
England: Griffin.
Bartholomew, D. J., Deary, I. J., & Lawn, M. (2009). A new lease of life for
Thomson's bonds model of intelligence. Psychol Rev, 116(3), 567-579.
Benson, N., Hulac, D. M., & Kranzler, J. H. (2010). Independent examination of the
Wechsler Adult Intelligence Scale-Fourth Edition (WAIS-IV): What does the
WAIS-IV measure? Psychological Assessment, 22(1), 121-130.
Bouchard, T. J. (1997). Experience producing drive theory: How genes drive
experience and shape personality. Acta Paediatrica, 86(S422), 60-64.
Brown, T. A. (2006). Confirmatory factor analysis for applied research. New York,
NY: Guilford Press; US.
Carroll, J. B. (1993). Human cognitive abilities: A survey of factor-analytic studies.
New York, NY: Cambridge University Press; US.
Cattell, R. B. (1943). The measurement of adult intelligence. Psychological bulletin,
40(3), 153.
Cattell, R. B. (1950). Personality. New York: McGraw-Hill.
Cattell, R. B. (1963). The theory of fluid and crystallized intelligence: A critical
experiment. Journal of Educational Psychology, 54, 1-22.
Cattell, R. B. (1987). Intelligence: Its structure, growth and action. New York:
Elsevier Science Publishing Company, Inc.
Cervone, D. (2005). Personality architecture: Within-person structures and processes.
Annual Reviews of Psychology, 56, 423-452.
Chamorro-Premuzic, T., & Furnham, A. (2006). Intellectual Competence and the
Intelligent Personality: A Third Way in Differential Psychology. Review of
General Psychology, 10(3), 251-267.
Chang, L., Connelly, B. S., & Geeza, A. A. (2012). Separating method factors and
higher order traits of the Big Five: A meta-analytic multitrait–multimethod
144
approach. Journal of Personality and Social Psychology; Journal of
Personality and Social Psychology, 102(2), 408.
Conn, S. R., & Rieke, M. L. (Eds.). (1994). 16PF Fifth Edition technical manual.
Champaign, IL: Institute for Personality and Ability Testing.
Costa Jr, P. T., & McCrae, R. R. (1992). NEO PI-R Professional Manual. Odessa,
FL: Psychological Assessment Resources.
Costa, P. T., & MacCrae, R. R. (1992). Revised NEO Personality Inventory (NEO
PI-R) and NEO Five-Factor Inventory (NEO FFI): Professional Manual:
Psychological Assessment Resources.
Costa, P. T., & McCrae, R. R. (1995). Primary traits of Eysenck's PEN system:
Three-and five-factor solutions. Journal of personality and social psychology,
69(2), 308.
Cramer, A. O., Sluis, S., Noordhof, A., Wichers, M., Geschwind, N., Aggen, S. H., . .
. Borsboom, D. (2012). Dimensions of normal personality as networks in
search of equilibrium: You can't like parties if you don't like people.
European Journal of Personality, 26(4), 414-431.
Cureton, E. E. (1968). A factor analysis of Project Talent tests and four other
batteries. Pittsburgh, PA: American Institutes for Research and School of
Education, University of Pittsburgh.
Dawis, R. V., & Lofquist, L. H. (1984). A psychological theory of work adjustment:
An individual-differences model and its applications: University of
Minnesota Press Minneapolis, MN.
De Fruyt, F., & Mervielde, I. (1999). RIASEC types and Big Five traits as predictors
of employment status and nature of employment. Personnel Psychology,
52(3), 701-727.
De Raad, B., & Barelds, D. P. (2008). A new taxonomy of Dutch personality traits
based on a comprehensive and unrestricted list of descriptors. Journal of
personality and social psychology, 94(2), 347.
De Raad, B., & Peabody, D. (2005). Cross‐culturally recurrent personality factors:
analyses of three factors. European Journal of Personality, 19(6), 451-474.
de Vries, R. E. (2011). No evidence for a General Factor of Personality in the
HEXACO Personality Inventory. Journal of Research in Personality, 45(2),
229-232.
Deary, I. J. (2009). The trait approach to personality G. M. Philip J. Corr (Ed.) The
Cambridge handbook of personality psychology (pp. 89-109).
Deng, C.-P., Armstrong, P. I., & Rounds, J. (2007). The fit of Holland’s RIASEC
model to US occupations. Journal of Vocational Behavior, 71(1), 1-22.
DeYoung, C. G. (2011). Intelligence and personality. In R. J. Sternberg & S. B.
Kaufman (Eds.), The Cambridge handbook of intelligence (pp. 711-737).
New York: Cambridge University Press.
DeYoung, C. G., Peterson, J. B., & Higgins, D. M. (2005). Sources of
openness/intellect: Cognitive and neuropsychological correlates of the fifth
factor of personality. Journal of personality, 73(4), 825-858.
145
DeYoung, C. G., Quilty, L. C., & Peterson, J. B. (2007). Between facets and
domains: 10 aspects of the Big Five. Journal of personality and social
psychology, 93(5), 880.
Dunkel, C. S. (2013). The general factor of personality and general intelligence:
Evidence for substantial association. Intelligence, 41(5), 423-427.
Einarsdóttir, S., & Rounds, J. (2000). Application of three dimensions of vocational
interests to the Strong Interest Inventory. Journal of Vocational Behavior,
56(3), 363-379.
Eysenck, H. J. (1939). Review of Thurstone's Primary Mental Abilities, 1938. British
Journal of Educational Psychology, 39, 188-231.
Eysenck, H. J. (1994). Personality and intelligence: Psychometric and experimental
approaches. In R. J. Sternberg & P. Ruzgis (Eds.), Personality and
intelligence (pp. 3-31). New York: Cambridge University Press.
Eysenck, H. J., & White, P. O. (1964). Personality and the measurement of
intelligence. British Journal of Educational Measurement, 24, 197-201.
Flanagan, D. P., Ortiz, S. O., & Alfonso, V. C. (2007). Essentials of cross-battery
assessment (2nd ed.). New Jersey: John Wiley & Sons, Inc.
Flanagan, J. C., Dailey, J. T., Shaycroft, M. F., Gorham, W. A., Orr, D. B., &
Goldberg, I. (1962). Design for a study of American youth. Boston: Houghton
Mifflin.
Flanagan, J. C., Davis, F. C., Dailey, J. T., Shaycroft, M. F., Orr, D. B., Goldberg, I.,
& Neyman, C. A. (1964). The American high-school student: The
identification, development, and utilization of human talents. Pittsburg:
Project Talent Office, University of PIttsburgh.
Goldberg, L. R. (1990). An alternative" description of personality": the big-five
factor structure. Journal of personality and social psychology, 59(6), 1216.
Gottfredson, L. S. (1986). Occupational aptitude patterns map: Development and
implications for a theory of job aptitude requirements. Journal of Vocational
Behavior, 29(2), 254-291.
Gottfredson, L. S. (2005). Applying Gottfredson's theory of circumscription and
compromise in career guidance and counseling. In S. D. Brown & R. W. Lent
(Eds.), Carreer development and counseling: Putting theory and research to
work (pp. 71-100). New York: John Wiley.
Guilford, J. P. (1959). Personality. New York: McGraw-Hill.
Gustafsson, J.-E. (1984). A unifying model for the structure of intellectual abilities.
Intelligence, 8(3), 179-203. doi: Doi: 10.1016/0160-2896(84)90008-4
Gustafsson, J.-E. (1988). Hierarchical models of individual difference in cognitive
abilities. In R. J. Sternberg (Ed.), Advances in the psychology of human
intelligence (Vol. 4). Hillsdale, New Jersey: Lawrence Erlbaum Associates,
Inc.
Gustafsson, J.-E. (2002). Measurement from a hierarchical point of view. In H. I.
Braun, D. N. Jackson & D. E. Wiley (Eds.), The role of constructs in
146
psychological and educational measurement (pp. 73-95). London: Lawrence
Erlbaum Associates, Publishers.
Hakistian, A. R., & Catell, R. B. (1975). The comprehensive ability battery.
Champaign, IL: Institute for Personality and Ability Testing.
Harring, J. R., Weiss, B. A., & Hsu, J. C. (2012). A comparison of methods for
estimating quadratic effects in nonlinear structural equation models.
Psychological Methods, 17(2), 193-214.
Hastie, T., & Tibshirani, R. (1986). Generalized additive models. Statistical science,
297-310.
Hogan, R. (1983). A socioanalytic theory of personality. In M. M. Page (Ed.),
Nebraska symposium on motivation: Vol. 30. Personality: Current theory and
research (pp. 55-89). Lincoln: University of Nebraska Press.
Hogan, R., & Roberts, B. W. (2000). A socioanalytic perspective on person–
environment interaction.
Holland, J. L. (1959). A theory of vocational choice. Journal of Counseling
Psychology, 6(1), 35.
Holland, J. L. (1973). Making vocational choices: A theory of careers (Vol. 37):
Prentice-Hall Englewood Cliffs, NJ.
Holland, J. L. (1985). Vocational Preference Inventory (VPI) manual: 1985 edition.
Odessa, FL: Psychological Assessment Resources.
Holland, J. L. (1997). Making vocational choices: A theory of vocational
personalities and work environments. Odessa, FL: Psychological Assessment
Resources
Hopwood, C. J., Wright, A. G. C., & Brent Donnellan, M. (2011). Evaluating the
evidence for the general factor of personality across multiple inventories.
Journal of Research in Personality, 45(5), 468-478. doi:
http://dx.doi.org/10.1016/j.jrp.2011.06.002
Horn, J. L. (1998). A basis for research on age differences in cognitive capacities. In
J. J. McArdle & R. W. Woodcock (Eds.), Human cognitive abilities in theory
and practice (pp. 29-73). Mahwah, NJ: Erlbaum.
Horn, J. L., & Blankson, N. (2005). Foundations for a better understanding of
cognitive abilities. In D. P. Flanagan & P. L. Harrison (Eds.), Contemporary
intellectual assessment: Theories, tests, and issues (pp. 41-68).
Horn, J. L., & Noll, J. (1997). Human cognitive abilities: Gf-Gc theory. In D. P.
Flanagan, J. L. Genshaft & P. L. Harrison (Eds.), Contemporary intellectual
assessment (pp. 53-91). New York, NY: Guilford Press.
Hu, L.-t., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance
structure analysis: Conventional criteria versus new alternatives. Structural
Equation Modeling: A Multidisciplinary Journal, 6(1), 1-55.
Humphreys, L. G., Lubinski, D., & Yao, G. (1993). Utility of predicting group
membership and the role of spatial visualization in becoming an engineer,
physical scientist, or artist. Journal of Applied Psychology, 78, 250-250.
Hunt, E. B. (2011). Human Intelligence. Cambridge: Cambridge University Press.
147
Irwing, P., Booth, T., Nyborg, H., & Rushton, J. P. (2012). Are g and the general
factor of personality (GFP) correlated? Intelligence, 40(3), 296-305.
Jackson, D. N. (1977). Manual for the Jackson Vocational Interest Survey. Port
Huron, MI: Research Psychologists Press.
Jensen, A. (1998). The g factor. Wesport, CN: Praeger.
Johnson, W., & Bouchard, T. J. (2009). Linking abilities, interests, and sex via latent
class analysis. Journal of Career Assessment, 17(1), 3-38.
Johnson, W., & Bouchard, T. J., Jr. (2005a). Constructive replication of the visual-
perceptual-image rotation model in Thurstone's (1941) battery of 60 tests of
mental ability. Intelligence, 33(4), 417-430.
Johnson, W., & Bouchard, T. J., Jr. (2005b). The structure of human intelligence: It
is verbal, perceptual, and image rotation (VPR), not fluid and crystallized.
Intelligence, 33(4), 393-416.
Johnson, W., Bouchard, T. J., Jr., McGue, M., Segal, N. L., Tellegen, A., Keyes, M.,
& Gottesman, I. I. (2007). Genetic and environmental influences on the
Verbal-Perceptual-Image Rotation (VPR) model of the structure of mental
abilities in the Minnesota study of twins reared apart. Intelligence, 35(6),
542-562.
Johnson, W., Jung, R. E., Colom, R., & Haier, R. J. (2008). Cognitive abilities
independent of IQ correlate with regional brain structure. Intelligence, 36(1),
18-28. doi: DOI: 10.1016/j.intell.2007.01.005
Johnson, W., Te Nijenhuis, J., & Bouchard, T. J., Jr. (2007). Replication of the
hierarchical visual-perceptual-image rotation model in de Wolff and Buiten's
(1963) battery of 46 tests of mental ability. Intelligence, 35(1), 69-81. doi:
http://dx.doi.org/10.1016/j.intell.2006.05.002
Judge, T. A., Higgins, C. A., Thoresen, C. J., & Barrick, M. R. (1999). The big five
personality traits, general mental ability, and career success across the life
span. Personnel Psychology, 52(3), 621-652.
Kan, K.-K., Kievet, R. A., Dolan, C. V., & van der Maas, H. L. (2011). On the
interpretation of the CHC factor Gc. Intelligence, 35(5), 292-302.
Kanfer, R., Wolf, M. B., Kantrowitz, T. M., & Ackerman, P. L. (2010). Ability and
trait complex predictors of academic and job performance: A person–
situation approach. Applied Psychology, 59(1), 40-69.
Kaufman, J. C. (2012). Self estimates of general, crystallized, and fluid intelligences
in an ethnically diverse population. Learning and Individual Differences, 22,
118-122.
Keith, T. Z., Fine, J. G., Taub, G. E., Reynolds, M. R., & Kranzler, J. H. (2006).
Higher order, multisample, confirmatory factor analysis of the Wechsler
Intelligence Scale for Children—Fourth Edition: What does it measure?
School Psychology Review, 35, 108-127.
Keith, T. Z., Reynolds, M. R., Patel, P. G., & Ridley, K. P. (2008). Sex differences in
latent cognitive abilities ages 6 to 59: Evidence from the Woodcock–Johnson
III tests of cognitive abilities. Intelligence, 36(6), 502-525.
148
Kenny, D. A. (2011). Measuring model fit Retrieved February 15, 2012, from
http://davidakenny.net/cm/fit.htm
Klein, A., & Moosbrugger, H. (2000). Maximum likelihood estimation of latent
interaction effects with the LMS method. Psychometrika, 65(4), 457-474.
Kruger, J., & Dunning, D. (1999). Unskilled and unaware of it: how difficulties in
recognizing one's own incompetence lead to inflated self-assessments.
Journal of Personality and Social Psychology; Journal of Personality and
Social Psychology, 77(6), 1121.
Kuder, G. F., & Richardson, M. W. (1937). The theory of test reliability.
Psychometrika, 2, 151-160.
Kuncel, N. R., Hezlett, S. A., & Ones, D. S. (2004). Academic performance, career
potential, creativity, and job performance: can one construct predict them all?
Journal of personality and social psychology, 86(1), 148-161.
Kvist, A. V., & Gustafsson, J.-E. (2008). The relation between fluid intelligence and
the general factor as a function of cultural background: A test of Cattell's
Investment theory. Intelligence, 36(5), 422-436. doi:
http://dx.doi.org/10.1016/j.intell.2007.08.004
Li, S.-C., Lindenberger, U., Hommel, B., Aschersleben, G., Prinz, W., & Baltes, P.
B. (2004). Transformations in the couplings among intellectual abilities and
constituent cognitive processes across the life span. Psychological Science,
15(3), 155-163. doi: 10.1111/j.0956-7976.2004.01503003.x
Loehlin, J. C. (2011). Correlation between general factors for personality and
cognitive skills in the National Merit twin sample. Journal of Research in
Personality, 45(5), 504-507.
Lohnes, P. R. (1966). Measuring adolescent personality: Project Talent five-year
follow-up studies, interim report 1. Pittsburgh, PA: University of Pittsburgh.
Lubinski, D. (2000). Scientific and social significance of assessing individual
differences:“Sinking shafts at a few critical points”. Annual review of
psychology, 51(1), 405-444.
Luciano, M., Leisser, R., Wright, M. J., & Martin, N. G. (2004). Personality, arousal
theory and the relationship to cognitive ability as measured by inspection
time and IQ. Personality and Individual Differences, 37(5), 1081-1089. doi:
http://dx.doi.org/10.1016/j.paid.2003.11.016
Lykken, D. T., Bouchard, T., McGue, M., & Tellegen, A. (1993). Heritability of
interests: a twin study. Journal of Applied Psychology, 78(4), 649.
Mackintosh, N. J. (2012). IQ and human intelligence (2nd ed.). New York, NY:
Oxford University Press.
Major, J. T., Johnson, W., & Deary, I. J. (2012). Comparing models of intelligence in
Project TALENT: The VPR model fits better than the CHC and extended Gf–
Gc models. Intelligence, 40(6), 543-559.
McCrae, R., & Costa, P. (1989). Reinterpreting the Myers-Briggs Type Indicator
from the perspective of the five-factor model of personality. Journal of
personality, 57(1), 17.
149
McCrae, R. R. (2009). The five-factor model of personality traits: Consensus and
controversy. In P. J. Corr & G. Matthews (Eds.), The Cambridge handbook of
personality psychology (pp. 148-161): Cambridge University Press.
McCrae, R. R., & Costa Jr, P. T. (2012). Personality in adulthood: A five-factor
theory perspective: Guilford Press.
McCrae, R. R., & Costa, P. T. (1985). Updating Norman's" adequacy taxonomy":
Intelligence and personality dimensions in natural language and in
questionnaires. Journal of personality and social psychology, 49(3), 710.
McCrae, R. R., & Terracciano, A. (2005). Universal features of personality traits
from the observer's perspective: data from 50 cultures. Journal of personality
and social psychology, 88(3), 547.
McGrew, K. S. (1997). Analysis of the major intelligence batteries according to a
proposed comprehensive Gf-Gc framework. In D. P. Flanagan, J. L. Genshaft
& P. Harrison (Eds.), Contemporary intellectual assessment: Theories, tests,
and issues (pp. 151-179). New York, NY: Guilford Press; US.
McGrew, K. S. (2005). The Cattell-Horn-Carroll Theory of Cognitive Abilities: Past,
present, and future. In D. P. Flanagan & P. L. Harrison (Eds.), Contemporary
intellectual assessment: Theories, tests, and issues (pp. 136-181). New York,
NY: Guilford Press; US.
McGrew, K. S. (2009). CHC theory and the human cognitive abilities project:
Standing on the shoulders of the giants of psychometric intelligence research.
Intelligence, 37(1), 1-10. doi: http://dx.doi.org/10.1016/j.intell.2008.08.004
McLarnon, M. J., & Carswell, J. J. (2012). The personality differentiation by
intelligence hypothesis: A measurement invariance investigation. Personality
and Individual Differences.
Moosbrugger, H., Schermelleh-Engel, K., Kelava, A., & Klein, A. G. (2009). Testing
multiple nonlinear effects in structural equation modeling: A comparison of
alternative estimation approaches. In T. Teo & M. S. Khine (Eds.), Structural
equation modelling in educational research: Concepts and applications.
Rotterdam, The Netherlands: Sense Publishers.
Moutafi, J., Furnham, A., & Paltiel, L. (2005). Can personality factors predict
intelligence? Personality and Individual Differences, 38, 1021-1033.
Myers, I. B., McCaulley, M. H., & Most, R. (1985). Manual: A guide to the
development and use of the Myers-Briggs Type Indicator: Consulting
Psychologists Press Palo Alto, CA.
Norman, W. T. (1963). Toward an adequate taxonomy of personality attributes:
Replicated factor structure in peer nomination personality ratings. The
Journal of Abnormal and Social Psychology, 66(6), 574.
O'Connor, B. (2000). SPSS and SAS programs for determining the number of
components using parallel analysis and Velicer's MAP test. Behavior
Research Methods, Instruments, & Computers, 32(2), 396-402.
Paulhus, D. L. (2002). Socially desirable responding: The evolution of a construct. In
H. I. Braun, D. N. Jackson & D. E. Wiley (Eds.), The role of constructs in
150
psychological and educational measurement (pp. 49-69). Mahwah, NJ:
Erlbaum.
Pincombe, J. L., Luciano, M., Martin, N. G., & Wright, M. J. (2007). Heritability of
NEO PI-R extraversion facets and their relationship with IQ. Twin Research
and Human Genetics, 10(3), 462-469.
Prediger, D. J. (1982). Dimensions underlying Holland's hexagon: Missing link
between interests and occupations? Journal of Vocational Behavior, 21(3),
259-287.
Raftery, A. E. (1995). Bayesian model selection in social research. Sociological
Methodology, 25, 111-163.
Raftery, A. E. (1995). Bayesian model selection in social research. Sociological
methodology, 25, 111-164.
Ramaswamy, V., DeSarbo, W. S., Reibstein, D. J., & Robinson, W. T. (1993). An
empirical pooling approach for estimating marketing mix elasticities with
PIMS data. Marketing Science, 12(1), 103-124.
Randahl, G. J. (1991). A typological analysis of the relations between measured
vocational interests and abilities. Journal of Vocational Behavior, 38(3), 333-
350.
Reeve, C. L. (2004). Differential ability antecedents of general and specific
dimensions of declarative knowledge: More than g. Intelligence, 32(6), 621-
652. doi: http://dx.doi.org/10.1016/j.intell.2004.07.006
Reeve, C. L., & Bonaccio, S. (2011). The Nature and Structure of “Intelligence”. The
Wiley-Blackwell Handbook of Individual Differences.
Reeve, C. L., & Hakel, M. D. (2000). Toward an understanding of adult intellectual
development: Investigating within-individual convergence of interest and
knowledge profiles. Journal of Applied Psychology, 85(6), 897.
Reeve, C. L., & Heggestad, E. D. (2004). Differential relations between general
cognitive ability and interest-vocation fit. Journal of Occupational and
Organizational Psychology, 77(3), 385-402. doi:
http://dx.doi.org/10.1348/0963179041752673
Reeve, C. L., Meyer, R. D., & Bonaccio, S. (2006). Intelligence-personality
associations reconsidered: The importance of distinguishing between general
and narrow dimensions of intelligence. Intelligence, 34(4), 387-402.
Rolfus, E. L., & Ackerman, P. L. (1996). Self-report knowledge: At the crossroads of
ability, interest, and personality. Journal of Educational Psychology, 88, 174-
188.
Rosenthal, R., & Rosnow, R. L. (1991). Essentials of behavioral research: Methods
and data analysis (Vol. 2): McGraw-Hill New York.
Rounds, J., & Tracey, T. J. (1993). Prediger's dimensional representation of
Holland's RIASEC circumplex. Journal of Applied Psychology, 78(6), 875.
Rushton, J. P., & Irwing, P. (2008). A General Factor of Personality (GFP) from two
meta-analyses of the Big Five: and. Personality and Individual Differences,
45(7), 679-683.
151
Sak, U. (2004). A synthesis of research on psychological types of gifted adolescents.
The Journal of Secondary Gifted Education, 15(2), 70-79.
Scarr, S. (1996). How people make their own environments: Implications for parents
and policy makers. Psychology, Public Policy, and Law, 2(2), 204.
Schmidt, F. L., & Hunter, J. (2004). General mental ability in the world of work:
Occupational attainment and job performance. Journal of personality and
social psychology, 86(1), 162-173.
Sclove, L. S. (1987). Application of model-selection criteria to some problems in
multivariate analysis. Psychometrika, 52, 333-343.
Shaycroft, M. F. (1967). The high school years: Growth in cognitive skills.
Pittsburgh, PA: American Institutes for Research and School of Education,
University of Pittsburgh.
Snow, R. E. (1963). Effects of learner characteristics in learning from instructional
films. Unpublished doctoral thesis, Purdue University.
Snow, R. E. (1989). Aptitude-treatment interaction as a framework for research on
individual differences in learning. In P. L. Ackerman, R. J. Sternberg & R.
Glaser (Eds.), Learning and individual differences: Advances in theory and
research (pp. 13-59). New York: Freeman.
Soubelet, A., & Salthouse, T. A. (2011). Personality-cognition relations across
adulthood. Developmental Psychology, 47(2), 303-310.
Spearman, C. (1904). " General Intelligence," Objectively Determined and
Measured. The American Journal of Psychology, 15(2), 201-292.
Spearman, C. (1939). Thurstone's work reworked. Journal of Educational
Psychology, 30(1-16).
Tabachnick, B. G., & Fidell, L. S. (2007). Using multivariate statistics (5th ed.).
New York, NY: Pearson.
Thurstone, L. L. (1938). Primary mental abilities. Psychometric Monographs, No.1
Chicago: University of Chicago Press.
Thurstone, L. L. (1947). Multiple factor analysis: A development and expansion of
The Vectors of the Mind. Chicago, IL: University of Chicago Press.
Tracey, T. J., & Rounds, J. (1992). Evaluating the RIASEC circumplex using high-
point codes. Journal of Vocational Behavior, 41(3), 295-311.
Tracey, T. J., & Rounds, J. (1996). The spherical representation of vocational
interests. Journal of Vocational Behavior, 48(1), 3-41.
Tracey, T. J., & Rounds, J. B. (1993). Evaluating Holland's and Gati's vocational-
interest models: A structural meta-analysis. Psychological bulletin, 113(2),
229.
Tupes, E. C., & Christal, R. E. (1992). Recurrent personality factors based on trait
ratings. Journal of personality, 60(2), 225-251.
Undheim, J. O. (1981). On intelligence IV: Toward a restoration of general
intelligence. Scandinavian Journal of Psychology, 22, 251-265.
152
Undheim, J. O., & Gustafsson, J.-E. (1987). The hierarchical organization of
cognitive abilities: Restoring general intelligence through the use of linear
structural relations (LISREL). Multivariate Behavioral Research, 22, 149-
171.
Van Der Maas, H. L., Dolan, C. V., Grasman, R. P., Wicherts, J. M., Huizenga, H.
M., & Raijmakers, M. E. (2006). A dynamical model of general intelligence:
The positive manifold of intelligence by mutualism. Psychological Review,
113(4), 842-861.
Van Iddekinge, C. H., Putka, D. J., & Campbell, J. P. (2011). Reconsidering
vocational interests for personnel selection: The validity of an interest-based
selection test in relation to job knowledge, job performance, and continuance
intentions. Journal of Applied Psychology, 96(1), 13.
Vernon, P. E. (1961). The structure of human abilities. London: Methuen.
Vernon, P. E. (1965). Ability factors and environmental influences. American
Psychologist, 20, 723-733.
Wai, J., Lubinski, D., & Benbow, C. P. (2009). Spatial ability for STEM domains:
Aligning over 50 years of cumulative psychological knowledge solidifies its
importance. Journal of Educational Psychology, 101(4), 817.
Waiyavutti, C., Johnson, W., & Deary, I. J. (2012). Do personality scale items
function differently in people with high and low IQ? Psychological
Assessment, 24(3), 545.
Wise, L. L., McLaughlin, D. H., & Steel, L. (1979). The Project TALENT data
hanbook, revised. Palo Alto, CA: American Institutes for Research.
Wolf, M. B., & Ackerman, P. L. (2005). Extraversion and intelligence: A meta-
analytic investigation. Personality and Individual Differences, 39(3), 531-
542.
Yang, C.-C. (2006). Evaluating latent class analysis models in qualitative phenotype
identification. Computational Statistics & Data Analysis, 50, 1090-1104.
Zeidner, M. (1995). Personality trait corrrelates of intelligence. In D. H. Saklofske &
M. Zeidner (Eds.), International handbook of personality and intelligence.
New York: Plenum Press.
Zeidner, M., & Shani-Zinovich, I. (2011). Do academically gifted and nongifted
students differ on the Big-Five and adaptive status? Some recent data and
conclusions. Personality and Individual Differences, 51(5), 566-570.
153
Appendix A: Supplemental tables for chapter 3
Table A1
First-order factor loadings for grade 10 males in the broad selection (VPR, CHC and Gf-Gc models)
Test Name Factor
Inform-
ation
English/
Math
Spatial/
Reasoning
Mech./
Science Speed Math
Vocabulary .64/.65/.65 .30/.31/.31
Literature .83/.84/.84
Music .73/.73/.73
Social Studies .84/.84/.84
Mathematics .32/.21/.21 .61/.68/.68
Physical Science .31/.32/.32 .47/.46/.46 .16/.15/.16
Biological Science .42/.43/.43 .36/.36/.36
Aeronautics and Space .33/.36/.36 .47/.47/.47
Electronics .82/.83/.83
Mechanics .74/.73/.73
Art .76/.76/.76
Law .68/.68/.68
Health .58/.59/.59 .05/.04/.04 .11/.13/.13
Bible .67/.67/.67
Theatre and Ballet .71/.71/.71
Miscellaneous .71/.71/.71
Memory for sentences .29/.29/.29
Memory for words .15/.14/.14 .39/.39/.39
Disguised words .33/.32/.32 .32/.34/.33 .25/.25/.26
Spelling .65/.64/.64
Capitalization .70/.69/.69
Punctuation .83/.83/.83
English usage .70/.70/.70
Effective expression .62/.62/.62
Word functions in sent. .42/.68/.67 .06/.09/.10 .33/–/–
Reading comprehension .53/.53/.54 .32/.31/.31 .10/.10/.10
Creativity .34/.36/.36 .16/.15/.15 .32/.30/.31
Mechanical reasoning .58/.57/.57 .33/.39/.39
Visualization in 2D .56/.56/.56 .23/.21/.20
Visualization in 3D .77/.78/.78
Abstract reasoning .32/.30/.30 .50/.51/.51
Math 1 .33/–/– .19/.16/.15 .37/.68/.68
Math 2 .13/–/– .78/.87/.97
Arithmetic comp. .26/–/– .34/.34/.34 .28/.50/.50
Table reading .72/.72/.73
Clerical checking .69/.69/.69
Object inspection .15/.15/.14 .59/.59/.59
Note. Loadings are in the order of VPR model, CHC model and Gf-Gc model
154
Table A2
First-order factor loadings for grade 10 females in the broad selection (VPR, CHC and Gf-Gc models)
Test Name Factor
Inform-
ation
English/
Math
Spatial/
Reasoning
Mech./
Science Speed Math
Vocabulary .58/.60/.60 .12/.13/.13 .23/.23/.23
Literature .83/.83/.83
Music .76/.76/.76
Social Studies .83/.83/.83
Mathematics .23/.12/.12 .67/.72/.72
Physical Science .66/.57/.56 .22/.31/.31
Biological Science .28/.35/.35 .45/.40/.40
Aeronautics and Space .16/.21/.21 .38/.35/.35
Electronics .59/.62/.62
Mechanics .17/.20/.20 .51/.50/.51
Art .77/.77/.77
Law .61/.61/.61
Health .47/.50/.50 .13/.10/.10 .11/.11/.11
Bible .62/.63/.63
Theatre and Ballet .73/.73/.73
Miscellaneous .67/.67/.67
Memory for sentences .39/.38/.38
Memory for words .20/.17/.17 .41/.44/.44
Disguised words .33/.31/.31 .37/.38/.38 .27/.27/.27
Spelling .66/.65/.66
Capitalization .70/.69/.69
Punctuation .85/.84/.85
English usage .69/.69/.69
Effective expression .09/.07/.07 .52/.53/.53
Word functions in sent. .43/.67/.66 .11/.14/.15 .33/–/–
Reading comprehension .53/.53/.53 .31/.30/.30 .11/.12/.12
Creativity .42/.42/.41 .34/.35/.35
Mechanical reasoning .64/.64/.64 .19/.19/.20
Visualization in 2D .59/.58/.58 .20/.20/.18
Visualization in 3D .75/.74/74
Abstract reasoning .34/.32/.32 .49/.50/.50
Math 1 .36/–/– .21/.20/.20 .31/.63/.63
Math 2 .21/–/– .69/.84/.84
Arithmetic comp. .44/–/– .29/.30/.31 .18/.57/.57
Table reading .25/.24/.23 .67/.67/.68
Clerical checking .69/.69/.68
Object inspection .24/.24/.22 .63/.63/.62
Note. Loadings are in the order of VPR model, CHC model and Gf-Gc model
155
Appendix B: Supplemental tables for chapter 4
Table B1
Standardized linear and quadratic effects of g on the raw personality scales (males)
Trait Linear effect Quadratic effect
Gr. 9 Gr. 10 Gr. 11 Gr.12 Gr. 9 Gr. 10 Gr. 11 Gr. 12
Sociability
Beta .155 .108 .066 .049 -.119 -.118 -.124 -.132
R2 .024 .012 .004 .002 .021 .020 .023 .025
Calmness
Beta .264 .261 .248 .246 – – – –
R2 .070 .068 .062 .061 – – – –
Vigor
Beta .233 .196 .165 .146 -.094 -.091 -.084 -.090
R2 .054 .038 .027 .021 .012 .011 .009 .011
Social Sensitivity
Beta .202 .202 .197 .193 – – -.018 -.024
R2 .041 .041 .039 .037 – – .000 .001
Tidiness
Beta .212 .188 .142 .114 -.058 -.052 -.061 -.068
R2 .045 .035 .020 .013 .004 .004 .005 .006
Culture
Beta .161 .156 .130 .132 – .023 .040 .048
R2 .026 .024 .017 .017 – .001 .003 .004
Self-Confidence
Beta .245 .218 .212 .222 – – – .017
R2 .060 .047 .045 .049 – – – .001
Mature Personality
Beta .273 .255 .241 .230 .044 .051 .054 .044
R2 .075 .065 .058 .053 .003 .005 .006 .004
Impulsiveness
Beta – .032 .030 .044 – – – –
R2 – .001 .001 .002 – – – –
Leadership
Beta .060 .071 .093 .116 .084 .080 .073 .065
R2 .004 .005 .009 .013 .010 .010 .008 .007
Effects greater than .015 are significant at p < .001, with no adjustment for multiple testing. Non-significant
effects are not shown.
156
Table B2
Standardized linear and quadratic effects of g on the personality scales (females)
Trait Linear effect Quadratic effect
Gr. 9 Gr. 10 Gr. 11 Gr.12 Gr. 9 Gr. 10 Gr. 11 Gr. 12
Sociability
Beta .129 .075 .028 -.023 -.138 -.149 -.148 -.138
R2 .017 .006 .001 .000 .030 .033 .033 .030
Calmness
Beta .226 .207 .201 .175 -.028 -.033 -.028 -.037
R2 .051 .043 .040 .031 .001 .001 .001 .002
Vigor
Beta .221 .177 .146 .118 -.078 -.088 -.077 -.067
R2 .049 .031 .021 .014 .008 .011 .009 .006
Social Sensitivity
Beta .237 .217 .209 .175 -.060 -.084 -.086 -.086
R2 .056 .047 .044 .031 .006 .011 .011 .011
Tidiness
Beta .176 .121 .086 .032 -.085 -.099 -.103 -.116
R2 .031 .015 .007 .001 .011 .014 .016 .020
Culture
Beta .254 .252 .247 .226 -.014 -.016 -.011 -.004
R2 .065 .063 .061 .051 .000 .000 .000 .000
Self-Confidence
Beta .185 .165 .164 .169 -.007 -.003 .011 .013
R2 .034 .027 .027 .029 .000 .000 .000 .000
Mature Personality
Beta .282 .264 .288 .269 .059 .055 .050 .036
R2 .080 .069 .083 .072 .005 .006 .004 .003
Impulsiveness
Beta .071 .119 .113 .117 .040 .046 .021 .023
R2 .005 .014 .013 .014 .002 .003 .001 .001
Leadership
Beta .074 .057 .072 .095 .052 .047 .050 .056
R2 .005 .003 .005 .009 .004 .004 .004 .005
Effects greater than .015 are significant at p < .001, with no adjustment for multiple testing. Non-significant
effects are not shown.
157
Table B3 AIC differences between null and full GAM models (grade 10 males/females) Trait AIC difference - males AIC difference - females Sociability 1451.7 1927.4 Calmness 371.5 13.2 Vigor 841.1 406.6 Social sensitivity 16.1 303.0 Tidiness 258.2 969.0 Culture 522.8 275.2 Self-confidence 502.8 116.8 Mature personality 805.8 1781.1 Impulsiveness 268.3 456.5 Leadership 1145.0 1132.3 Note: the AIC for null model in males was 134924.2, and 135009.3 in females.
158
Appendix C: Supplemental tables for chapter 5
Table C1. Odds ratios of abilities and interests predicting job categories (grade 11 males).
Predictor Job Category
Science Med. Business Teach. Humanities Fine Arts Technical Sales Mech. Clerical Construc. Labour
g 2.64 2.97 1.38 1.34 3.50 1.32 1.33 .56 .66 .50
Spatial 1.38 1.35 .73 1.33 1.26
English 1.43 1.19 .79 1.25
Speed 1.23 .77 1.24
Math 1.39 1.33 .84 1.17
I: Trades .73 .70 .81 .78 .46 1.65 1.49
I: Politics 1.29 .72 1.43 .78
I: Science 1.60 1.56 .78 .81 1.31 1.15 1.50
I: Clerical 1.15 1.16 .82 .65 1.27 .82
I: Medicine 2.27 .82 .77
I: Arts .75 1.21 .82 1.55 1.44 .70 1.17 .81 1.18
I: Teaching 1.74 1.35 .76 .70 1.27
I: Military 1.25 .77 1.19
I: Sales .61 .77 1.18 1.66 1.71 .68 .78 1.54
I: Arch. 1.54 .57 1.22 .79 1.29 1.28 .82
F1: Trades .17 .11 .55 .70 .08 .55 .81 3.05 1.71 3.71 2.91
F2: People .23 1.16 2.66 2.55 .58 .56 .48 1.70 .76
F3: Things 19.58 5.67 1.56 .61 2.62 3.70 2.31 .82 .80 .40 .53 .34
R2: full .38 .49 .10 .20 .51 .20 .09 .03 .24 .13 .23 .02
R2: factors .53 .51 .09 .18 .61 .17 .09 .01 .28 .10 .27 .21
Note: Odds ratios between .86 and 1.14 not shown. R2: full was for individual scales as predictors; R
2: factors was for the trait-complex factors.
Med. = Medicine, Mech. = Mechanical.
159
Table C2. Odds ratios of abilities and interests predicting job categories (grade 11 females).
Predictor Job Category
Science Med. Business Teaching Humanities Fine
Arts
Technical Sales Mech. Clerical Construc. Labour
g 2.51 1.96 1.83 1.33 1.82 .64 .80 .32 .28 .52
Spatial 1.40 .72 .79 .83 1.41 1.33 1.90 .39
English 1.70 .67 .83 .80 1.34 .80 1.40 .40
Speed .84 1.33 .56 1.63 .56
Math .79 .78 1.17 1.34
I: Trades .34 1.26 .82 .85 .55 1.61 1.27
I: Politics 1.42 1.20 1.37 1.21 .82 1.49 .62 .65
I: Science 2.47 1.24 .84 .58 1.39 1.77
I: Clerical 2.47 1.33 .49 .69 .47 1.25 1.41 .43
I: Medicine 2.75 3.23 .76 .82 1.90 1.15 .79 1.62
I: Arts .82 1.37 1.53 1.15 1.23 1.25 1.88 1.50
I: Teaching 1.67 .81 1.70 1.25 2.35 .82 .56 .34 .81
I: Military 1.44 .75 16.63 1.33
I: Sales .34 .64 1.15 .65 1.33 .66 1.66 .84
F1: Clerical .53 .10 .58 .29 .32 .46 1.18 2.56 1.52 1.26 1.90
F2: People 1.24 2.66 5.20 2.59 2.97 .85 1.22 .27 .68 .44 .40
F3: Things 1.74 5.67 .51 .52 .63 .83 1.56 1.25 4.06 1.77
R2: full .53 .36 .19 .25 .26 .30 .19 .13 .45 .06 .80 .18
R2: factors .18 .51 .11 .32 .29 .16 .01 .02 .24 .04 .22 .11
Note: Odds ratios between .86 and 1.14 not shown. R2: full was for individual scales as predictors; R
2: factors was for the trait-complex factors.
Med. = Medicine, Mech. = Mechanical.
160
Table C3. Odds ratios of latent classes in predicting job category (grade 11 males).
Predictor Job Category
Science Med. Business Teaching Humanities Fine
Arts
Technical Sales Mech. Clerical Construc. Labour
Class 1 .56 .09 .47 .79 .68 1.17 1.95 1.14 1.92
Class 2 .83 .41 .70 .22 .79 1.17 1.51 .80
Class 3 .15 .76 3.24 3.66 1.26 .18 .80 2.03 .32 1.43
Class 5 .60 .37 1.25 2.29 2.21 1.18 1.21 .62 1.24 1.47 .75
R2 .10 .13 .03 .06 .22 .01 .01 .10 .02 .02 .07 .01
Note: Reference class is class 4. Odds ratios between .86 and 1.14 not shown. Med. = Medicine, Mech. = Mechanical
Table C4. Odds ratios of latent classes in predicting job category (grade 11 females).
Predictor Job Category
Science Med. Business Teaching Humanities Fine
Arts
Technical Sales Mech. Clerical Construc. Labour
Class 1 .11 .35 .67 .73 .53 1.53 n/a 1.21
Class 3 1.70 .73 2.11 1.89 1.52 1.27 1.30 .51 1.20 n/a .73
Class 4 .58 2.60 1.24 2.20 5.29 1.57 1.29 1.48 .19 .67 n/a .54
Class 5 .26 1.33 .42 2.67 1.24 .52 .60 .73 1.27 n/a 1.35
Class 6 .72 3.00 .85 2.00 .70 2.90 .32 1.49 .24 n/a 1.64
R2 .23 .12 .04 .05 .15 .02 .04 .02 .09 .01 n/a .03
Note: Reference class is class 2. Odds ratios between .86 and 1.14 not shown. Med. = Medicine, Mech. = Mechanical
a model did not converge.